w-xb/awesome-agentic-robotics
w-xb/awesome-agentic-robotics has 31 GitHub stars and 6 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-16 04:29 UTC. Last push 2026-11-07.
Measured star growth
- in 24h
- 0
- in 7 days
- Update delayed
- in 15 days
- Update delayed
- in 1 month
- Update delayed
Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.
About
A curated survey of embodied agents:robotic memory, planning, world models, verification, and failure recovery
LearningREADME.md
OriginalAwesome Agentic Robotics 
Agentic Robotics studies autonomous physical agents that reason, remember, plan, use tools, verify execution, recover from failure, and improve from embodied experience.
A selective, living collection of influential papers, systems, benchmarks, and open-source resources spanning agent orchestration, VLA, memory, world models, verification, recovery, self-evolution, and physical safety.
Last updated: 2026-09-11
Contents
- What Agentic Robotics Means Here
- Taxonomy: Agent–Harness–VLA Stack
- Scope and Curation
- Start Here / Must Read
- Surveys and Position Papers
- Agentic Robotics Architectures
- Agentic VLA and Embodied Foundation Models
- Self-Evolving Embodied Agents
- Embodied Memory
- Planning and Reasoning
- World Models and World-Action Models
- Verification and Safety Evaluation
- Failure Detection and Recovery
- Skill Learning and Tool Use
- Long-Horizon Manipulation
- Embodied Navigation
- Human-Robot Interaction
- Governance and Physical Safety
- Benchmarks and Datasets
- Foundations and Industrial Systems
- Extended Reading
- Related Awesome Lists
- Citation
What Agentic Robotics Means Here
Agentic Robotics is an emerging and not-yet-standardized research direction. In this repository, we use an ability-centered definition: a robotic system becomes more agentic as it can perceive and maintain task state, use memory, plan over future consequences, invoke skills or tools, verify execution, recover from failures, adapt from experience, and act under physical safety constraints.
Unlike many software agents, whose planning, memory, tool use, and reflection are mediated mainly through text and APIs, robotic agents must act in a dynamic, partially observable physical world. Their agency therefore depends not only on an external LLM-style loop, but also on embodied mechanisms such as spatial-temporal memory, action-conditioned world models, grounded skill interfaces, execution monitoring, failure attribution, recovery policies, and safety envelopes.
Agentic robotics is a continuum rather than a binary label. Not every work in this list presents a complete autonomous robot: some contribute a single capability, while others integrate several capabilities into a closed-loop system. This repository maps both the components of physical agency and the architectures that bring them together.
Taxonomy: Agent–Harness–VLA Stack
The organizing view behind this repository is a closed-loop stack:
Goal / Human Intent
↓
Agent / Planner
(reasoning · decomposition · task state)
↓
Harness / Runtime
(memory · tools · routing · verification · recovery · safety)
↓
VLA / Policy
(grounded perception-to-action execution)
↓
Robot + Physical World
│
└── observations · outcomes · failures ──↺
Persistent, validated updates across episodes → Self-Evolution
The agent decides what should happen; the harness coordinates components and governs execution; the VLA/policy performs grounded physical actions; the physical world supplies evidence about success, failure, and safety. Self-evolution begins when validated changes persist and influence future episodes.
This stack is a research lens rather than a claim that every system must use the same modular architecture. Papers are placed by their primary agentic contribution, while secondary capabilities are described in the entry. See TAXONOMY.md for layer boundaries, evolution targets, and open research questions.
Scope and Curation
Core scope. The list covers planning and grounded reasoning; VLA-as-tools and hierarchical VLA; self-evolving skills, tools, policies, objectives, verifiers, data pipelines, and runtime harnesses; embodied memory and persistent state; world models used for prediction, evaluation, or control; execution verification and recovery; skill learning and robot tool use; long-horizon manipulation and navigation; human-robot interaction; and governance, runtime safety, and embodied evaluation.
Selection principles. Entries are chosen for agentic relevance, influence, evidence quality, representativeness, and resource completeness. The list includes both modular LLM/VLM-agent architectures and tightly coupled robot-learning systems when they make a clear contribution to closed-loop physical agency. General-purpose VLA work is included only when it materially advances planning, long-horizon execution, tool use, verification, recovery, embodied memory, world-model-based decisions, continual adaptation, or cross-embodiment deployment.
A living list. New high-quality work may be added as the field develops; selectivity is maintained through a quality bar, not a fixed item count. Each paper has one canonical entry, items are ordered by first public release date within each section, and narrower or superseded work remains available in Extended Reading.
Canonical placement. A paper is categorized by the component or capability it changes most directly. Secondary capabilities are reflected in the description rather than by duplicating the paper across sections.
Links: Each canonical title links to the official publication or preprint. Supplementary resources use [Project], [Code], [Model], and [Dataset].
Start Here / Must Read
Twenty-five high-signal papers organized as five reading routes. Each also appears once in its canonical topic section.
VLA as Tools and Agents
Harness VLA — Treats a frozen VLA as a retryable contact-rich primitive, while a memory-guided agent stages subtasks, models failures, verifies outcomes, and retries execution. It is a strong blueprint for composing existing policies into reliable long-horizon systems.
VLAs-as-Tools — Separates temporal reasoning from physical execution through a high-level VLM agent and specialized, progress-aware VLA tools. Event-triggered replanning and tool-aligned post-training make the architecture especially useful for understanding modular agentic manipulation.
Agentic Robot — Introduces a Plan-Execute-Verify loop that combines task planning, VLA execution, temporal verification, and structured recovery. The explicit closed loop makes it a representative system for studying how embodied agents detect and repair execution errors.
ART — Injects off-the-shelf multimodal tools into an end-to-end VLA without abandoning continuous action generation. Tool-use trajectories and long-horizon reasoning training demonstrate how modular capabilities can improve robustness, data efficiency, and extensibility in simulation and real robots.
Gemini Robotics 1.5 — Couples a generalist robot policy with embodied reasoning, explicit thinking, and motion transfer across embodiments. It is an important industrial-scale example of moving beyond direct instruction following toward generalizable, reasoning-guided physical action.
Self-Evolution Highlights
Zetta ζ — Freezes the base VLA and high-level orchestrator while evolving runtime critics, recovery playbooks, and typed tools. Its action-frequency monitoring, failure-cluster diagnosis, replay, same-seed regression, and held-out promotion make it the clearest current reference for versioned harness evolution in embodied execution.
SHAPER — Separately evolves reusable textual skills and a context-code harness around frozen planner and executor models. Its two-stage skill-then-harness search, sandboxed candidates, and top-k validation provide a useful contrast to Zetta's action-frequency runtime governance.
ASPIRE — Turns robot execution traces into causal failure diagnoses, repaired programs, and reusable skills through iterative exploration. The distinction between one-off repair and validated skill abstraction makes it a strong reference for failure-driven skill evolution.
VASO — Refines embodied skill contracts against temporal specifications and model-checking counterexamples. By making preconditions, local rules, and planner-facing semantics explicit, it connects self-evolving skills to formal verification and physical safety constraints without updating foundation-model weights.
ENPIRE — Runs a coding-agent loop on real robots that edits policy-improvement algorithms, rewards, training infrastructure, and evaluation code from rollout evidence. It represents a deeper policy and training evolution layer than Zetta's frozen-policy runtime harness.
Hierarchical and Reasoning VLA
τ_0-VLA — Combines hierarchical control, execution memory, a learned world model, and test-time search to improve high-level subtask decisions. It shows how additional inference-time computation can be converted into more reliable long-horizon robot behavior.
G0.5 — Generates reasoning and action tokens in one autoregressive stream, supported by visual memory and cross-embodiment action tokenization. The unified formulation is a useful reference for tightly integrating deliberation and control inside a foundation model.
Qwen-VLA — Unifies manipulation, navigation, trajectory supervision, multimodal understanding, and cross-embodiment training within one model family. Its broad task coverage makes it a valuable reference for studying generalist embodied foundations rather than isolated robot skills.
π0.5 — Extends a generalist VLA with heterogeneous co-training and high-level semantic knowledge for open-world generalization. It is a key reference for understanding how web-scale knowledge and robot data can support deployment in unseen environments.
Hi Robot — Uses a hierarchical VLA to convert open-ended human instructions into semantic subgoals and low-level actions. It clearly illustrates why task-level reasoning and policy-level execution benefit from distinct but jointly coordinated representations.
Memory and World Models
AtlasVLA — Maintains a persistent world-ego state that tracks spatial context and task progress across long executions. It directly addresses perceptual and progress forgetting, two central obstacles to reliable memory-aware VLA control.
MEMORA — Extracts editable, consolidated embodied action memories from egocentric experience and retrieves them for later reasoning and planning. It offers a concrete path from passive video experience to reusable procedural knowledge for physical agents.
MemoryVLA — Integrates perceptual and cognitive memory into VLA control so the policy can preserve scene evidence and task state over time. It is a representative memory architecture for long-horizon manipulation under partial observability.
τ_0-WM — Unifies action generation, future video prediction, candidate evaluation, and simulator-based rectification in one video-action world model. The design makes future imagination operational by using predictions to select and correct robot actions.
V-JEPA 2 — Learns predictive video representations at scale and demonstrates understanding, anticipation, and planning on physical tasks. It is a foundational reference for how self-supervised world models can support downstream embodied decision making.
Planning, Safety, and Evaluation
SayCan — Grounds language-model plans in learned robotic affordances, scoring candidate skills by both semantic usefulness and physical feasibility. This influential formulation established a practical pattern for connecting broad language knowledge to executable robot capabilities.
Inner Monologue — Adds environment feedback, success detection, and replanning to language-model-based control, converting open-loop plans into an embodied reasoning loop. It remains a foundational reference for feedback-aware planning and recovery in robotics.
MANIGUARD — Evaluates manipulation safety against explicit specifications and pairs the benchmark with safety-annotated trajectories for improvement. It is valuable because it measures process-level behavior rather than relying only on final task success.
RoboCasa365 — Provides a large-scale, diverse simulation framework for training and evaluating generalist robot policies over everyday tasks. Its breadth makes it a central testbed for long-horizon generalization and agentic policy orchestration.
Open X-Embodiment — Aggregates large-scale robot datasets across institutions and embodiments and introduces RT-X models trained on the mixture. It established shared data and cross-robot transfer as core foundations for generalist embodied intelligence.
Surveys and Position Papers
Field definitions, surveys, and position papers that establish the concepts, taxonomy, capabilities, risks, and open questions of Agentic Robotics.
- 2026.7 · Self-Improvements in Modern Agentic Systems: A Survey - Defines agents as foundation models coupled to operational scaffolds and organizes self-improvement by update target, driving signal, evaluation, and open problems Resources: Project · Code.
- 2026.5 · World Action Models: The Next Frontier in Embodied AI - Defines world-action modeling, organizes its major architectural choices, and identifies how predictive environment dynamics can improve planning and closed-loop robot control.
- 2026.5 · World Model for Robot Learning: A Comprehensive Survey - Reviews world-model objectives, representations, datasets, and downstream uses across robot learning, providing a broad foundation for prediction-guided embodied decision making.
- 2026.4 · Vision-Language-Action Safety: Threats, Challenges, Evaluations, and Mechanisms - Systematizes VLA safety threats, evaluation protocols, and mitigation mechanisms across training and deployment, connecting model risks to physical execution consequences.
- 2025.8 · Towards Embodied Agentic AI: Review and Classification of LLM- and VLM-Driven Robot Autonomy and Interaction - Classifies language- and vision-driven robot agents by autonomy, interaction, architecture, and embodiment, clarifying the emerging design space of agentic robotics.
- 2025.7 · A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence - Organizes self-evolution across model, context, tools, and agentic architecture, distinguishing transient test-time adaptation from persistent changes that affect future tasks.
- 2023.11 · Large Language Models for Robotics: A Survey - Surveys early uses of large language models for perception, planning, control, and human-robot interaction, establishing essential context for modern agentic systems.
Agentic Robotics Architectures
End-to-end or multi-module robot-agent systems that organize planning, execution, memory, verification, recovery, skill selection, tool interfaces, or multi-agent coordination.
- 2026.8 · ETA: A New Agentic Paradigm for Embodied Tasks - Presents an agentic embodied-task framework that coordinates reasoning and execution through a minimal interface, emphasizing transferable orchestration over monolithic policy scaling.
- 2026.7 · Embodied Agents Take Control: Minimal-Interface Zero-Shot Agents Rival Industrial-Scale Policies in Vision-and-Language Navigation - Shows that minimally interfaced foundation-model agents can compete with specialized navigation policies through zero-shot reasoning, feedback, and structured action selection.
- 2026.6 · HoloAgent-0: A Unified Embodied Agent Framework with 3D Spatial Memory - Integrates 3D spatial memory, reasoning, and action in a unified embodied framework for persistent scene understanding and long-horizon task execution Resources: Project · Code.
- 2026.6 · What Matters in Orchestrating Robot Policies: A Systematic Study of Hierarchical VLA Agents - Systematically isolates orchestration choices in hierarchical VLA agents, revealing which planning, policy-selection, and feedback mechanisms most affect long-horizon performance Resources: Project.
- 2026.6 · VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation - Coordinates perception, open-vocabulary reasoning, and reusable manipulation policies through a physical orchestrator designed for compositional long-horizon tasks in changing scenes Resources: Project.
- 2026.4 · A Physical Agentic Loop for Language-Guided Grasping with Execution-State Monitoring - Closes the grasping loop with execution-state monitoring, allowing a language-guided agent to detect physical progress and adapt its next action Resources: Project · Code.
- 2023.9 · SMART-LLM: Smart Multi-Agent Robot Task Planning using Large Language Models - Uses language models to decompose tasks, allocate roles, and coordinate heterogeneous robots, offering an influential architecture for multi-agent embodied planning Resources: Project · Code.
- 2022.4 · Do As I Can, Not As I Say: Grounding Language in Robotic Affordances - Grounds language-model planning in learned affordance values, ensuring selected skills are both semantically appropriate and physically executable by the robot Resources: Project.
Agentic VLA and Embodied Foundation Models
VLA-centric systems that expose robot policies as skills or tools, add reasoning, verification, recovery, memory, or online adaptation, or provide hierarchical and cross-embodiment foundations for agentic control.
- 2026.8 · τ_0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation - Uses execution memory and world-model-guided test-time search to improve high-level subtask decisions in a hierarchical robot foundation model.
- 2026.8 · Evolve Vision-Language-Action Model into an Agent with On-the-fly Tool-use - Introduces ART, which injects dynamic multimodal tool calls into VLA control and trains long-trajectory tool reasoning for robust physical execution.
- 2026.8 · G0.5: One Autoregressive Stream for Robot Reasoning and Action - Produces interleaved reasoning and actions with visual memory and cross-embodiment tokenization, unifying deliberation and robot control in one stream.
- 2026.7 · Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents - Wraps frozen VLAs as retryable manipulation primitives and adds task memory, failure modeling, verification, and agent-controlled recovery for long-horizon execution Resources: Project · Code.
- 2026.5 · Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments - Unifies manipulation, navigation, multimodal understanding, and trajectory learning across environments and robot embodiments within a broadly trained foundation model.
- 2026.5 · Agentic-VLA: Efficient Online Adaptation for Vision-Language-Action Models - Adds reward generation, language-guided exploration, and experience-weighted memory around a VLA, enabling efficient online adaptation to new manipulation tasks.
- 2026.5 · Towards Long-horizon Embodied Agents with Tool-Aligned Vision-Language-Action Models - Frames specialized VLAs as tools for a high-level agent, with progress feedback, event-triggered replanning, and tool-aligned post-training for long horizons.
- 2025.10 · Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer - Combines generalist control, embodied reasoning, explicit thinking, and motion transfer to support complex tasks across multiple robot embodiments.
- 2025.5 · Agentic Robot: A Brain-Inspired Framework for Vision-Language-Action Models in Embodied Agents - Coordinates planning, VLA execution, temporal verification, and error recovery through a structured Plan-Execute-Verify loop for embodied agents and autonomous recovery Resources: Project · Code.
- 2025.4 · π0.5: a Vision-Language-Action Model with Open-World Generalization - Combines heterogeneous robot data and semantic knowledge to improve open-world generalization for manipulation in unseen environments and tasks Resources: Project · Code.
- 2025.2 · Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models - Decomposes open-ended instructions into semantic subgoals and low-level VLA actions, demonstrating hierarchical reasoning for interactive robot control across diverse tasks Resources: Project.
Self-Evolving Embodied Agents
Systems that use execution or training feedback to propose, validate, and retain changes to a robot agent's skills, tools, policy, objective, memory, verifier, or runtime harness. Transient replanning, test-time adaptation, and fixed-library recovery remain in their canonical or extended sections.
Layer key. harness/runtime · skill/library · contract/verifier · policy/data · objective/reward · memory/context. Digital, transient, and narrower-layer precedents are collected in Extended Reading.
- 2026.9 · Recursive Harness Distillation across Agents for Robot Manipulation - Memory/context. Distills Astra's robotic intervention experience into a reusable playbook for Luna and recursively refines it using Luna's execution feedback, keeping agent and VLA model weights fixed. Resources: Project.
- 2026.8 · Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence - Harness/runtime. Freezes a base VLA and high-level orchestrator while evolving runtime critics, recovery playbooks, and schema-constrained tools through failure diagnosis, replay, regression gates, and held-out promotion Resources: Project · Code.
- 2026.8 · Self-Evolving Embodied Agents via Skill-Harness Evolution - Skill + harness. Evolves reusable textual skills and a context-code harness around frozen planner and executor models using episode-level feedback, beam search, sandboxed candidates, and top-k validation.
- 2026.6 · ASPIRE: Agentic /Skills Discovery for Robotics - Skill/library. Diagnoses robot-program failures from execution traces, synthesizes repairs, and consolidates validated programs into reusable skills through iterative agentic exploration Resources: Project · Code.
- 2026.6 · ENPIRE: Agentic Robot Policy Self-Improvement in the Real World - Policy/training code. Runs a coding-agent loop on real robots that edits policy-improvement algorithms, rewards, training infrastructure, and evaluation code from rollout evidence.
- 2026.6 · VASO: Formally Verifiable Self-Evolving Skills for Physical AI Agents - Contract/verifier. Refines embodied skill contracts against temporal specifications and model-checking counterexamples, connecting self-evolving skills to formal verification and physical safety constraints Resources: Project.
- 2026.5 · EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents - Skill/context. Maintains persistent textual skills and classifies reflections as discovery, optimization, skill defects, or execution lapses before targeted consolidation edits.
- 2026.5 · Learning While Deploying: Fleet-Scale Reinforcement Learning for Generalist Robot Policies - Policy/data (human-assisted). Closes an offline-to-online fleet loop in which shared autonomous rollouts and human interventions update a generalist VLA policy across sixteen real robots.
- 2026.3 · Uni-Skill: Building Self-Evolving Skill Repository for Generalizable Robotic Manipulation - Skill/library. Builds a reusable skill repository that retrieves, adapts, and consolidates manipulation capabilities from experience, making skill-library growth an explicit learning target.
- 2026.2 · RISE: Self-Improving Robot Policy with Compositional World Model - Policy/data. Uses a compositional world model and progress-aware value estimates to generate improvement data and update robot policies through imagined and real trajectories Resources: Project.
- 2024.5 · Agentic Skill Discovery - Skill/objective (simulation). Combines language-model task and reward generation, reinforcement learning, and independent visual verification to discover robot skills and add them to a persistent library Resources: Project.
- 2023.10 · Eureka: Human-Level Reward Design via Coding Large Language Models - Objective/reward (simulation). Searches executable reward programs with language-model proposals and simulator feedback, showing objective-level self-improvement for robot skill policies Resources: Project.
- 2023.6 · RoboCat: A Self-Improving Generalist Agent for Robotic Manipulation - Policy/data. Combines demonstrations with autonomous rollouts to collect data on new tasks and fine-tune a generalist manipulation policy, establishing an early real-robot example of continual policy self-improvement.
Embodied Memory
Memory systems for physical agents, including working and episodic memory, semantic-spatial scene representations, experience retrieval, and persistent world or task state.
- 2026.8 · AtlasVLA: Persistent World-Ego State Modeling for Vision-Language-Action Models - Maintains a persistent world-ego state that preserves spatial context and task progress, reducing forgetting during long-horizon VLA execution.
- 2026.7 · MEMORA: Embodied Action Memory from Egocentric Videos for Reasoning and Planning - Converts egocentric experience into editable, consolidated action memories that can be retrieved and recomposed for later embodied reasoning and planning Resources: Project.
- 2026.3 · Chameleon: Episodic Memory for Long-Horizon Robotic Manipulation - Uses episodic experience retrieval to adapt manipulation behavior over long horizons, connecting remembered outcomes to current task decisions.
- 2026.3 · MEM: Multi-Scale Embodied Memory for Vision Language Action Models - Builds multi-scale memory over observations and actions so VLA policies can retain both short-term detail and long-term task context Resources: Project.
- 2025.10 · MemER: Scaling Up Memory for Robot Control via Experience Retrieval - Scales experience retrieval for robot control, selecting relevant prior trajectories to improve action decisions under new tasks and environments Resources: Project · Code.
- 2025.8 · MemoryVLA: Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation - Integrates perceptual and cognitive memory into VLA control, preserving scene evidence and task state for long-horizon manipulation under partial observability Resources: Project · Code.
- [2024.11 · [CVPR 2025] 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning](https://arxiv.org/abs/2411.17735) - Constructs an explicit 3D scene memory that supports persistent exploration, spatial grounding, and reasoning across changing embodied viewpoints Resources: Project · Code.
- 2024.11 · DynaMem: Online Dynamic Spatio-Semantic Memory for Open World Mobile Manipulation - Updates spatio-semantic memory online as environments change, enabling open-world mobile manipulation with persistent object and scene knowledge during extended deployments Resources: Project · Code.
Planning and Reasoning
LLM/VLM-based task planning, grounded reasoning, decomposition, formal or programmatic plans, value maps, future-state prediction, and closed-loop replanning from feedback.
- 2025.12 · Large Video Planner - Plans long-horizon robot behavior through generated visual futures, using video-level predictions to represent intermediate physical states and actions Resources: Project · Code.
- 2025.9 · Reinforced Embodied Planning with Verifiable Reward for Real-World Robotic Manipulation - Trains embodied planning with automatically verifiable rewards, improving structured reasoning and real-world manipulation without relying solely on imitation traces.
- 2025.7 · ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning - Reinforces planning in a visual latent space before action generation, improving VLA reasoning while avoiding expensive natural-language chain-of-thought supervision Resources: Project.
- 2025.3 · Embodied-Reasoner: Synergizing Visual Search, Reasoning, and Action for Embodied Interactive Tasks - Jointly trains visual search, reasoning, and action selection so embodied agents can gather missing evidence before committing to task decisions Resources: Project.
- 2025.3 · Code-as-Symbolic-Planner: Foundation Model-Based Robot Planning via Symbolic Code Generation - Generates executable symbolic planning code from multimodal task context, combining foundation-model understanding with structured constraints and reusable planning abstractions Resources: Project.
- 2024.2 · Grounding LLMs For Robot Task Planning Using Closed-loop State Feedback - Grounds language-model plans in observed state feedback, allowing task steps to be revised when physical execution changes the environment unexpectedly.
- 2024.2 · Verifiably Following Complex Robot Instructions with Foundation Models - Translates complex language instructions into formal task specifications and verifies execution, connecting foundation-model interpretation with reliable robot planning.
- 2023.7 · VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models - Converts language and perception into composable 3D value maps, enabling zero-shot spatial reasoning and closed-loop manipulation planning from language instructions Resources: Project.
- 2022.12 · LLM-Planner: Few-Shot Grounded Planning for Embodied Agents with Large Language Models - Uses few-shot language-model prompting and environmental grounding to generate executable plans for embodied tasks with limited demonstrations or domain-specific training.
- 2022.9 · ProgPrompt: Generating Situated Robot Task Plans using Large Language Models - Represents robot plans as situated programs with assertions and recovery steps, improving executability and robustness under changing environments.
- 2022.7 · Inner Monologue: Embodied Reasoning through Planning with Language Models - Feeds success signals, scene descriptions, and human feedback back into language-model planning, enabling closed-loop embodied reasoning and replanning.
- 2022.1 · Language Models as Zero-Shot Planners: Extracting Actionable Knowledge for Embodied Agents - Extracts actionable task sequences from pretrained language models and grounds them to available actions, establishing a foundational zero-shot planning baseline.
World Models and World-Action Models
Predictive models that learn environment dynamics or joint video-action structure to support future imagination, candidate evaluation, control, action synthesis, and internal planning.
- 2026.8 · World Tokens: Enhancing Embodied Policies with Training-Time World Modeling - Adds predictive world tokens during training to improve embodied policies while avoiding the deployment cost of explicit future generation.
- 2026.8 · Faster-WAM: Efficient Inference-Time Future Conditioning for Robust World Action Models - Retains future-conditioned action selection while reducing video-action interaction cost, making inference-time imagination more practical for robust control during online execution.
- 2026.7 · ST-WAM: Semantic-Temporal World Action Model for Robust Manipulation under Visual Distribution Shifts - Combines DINOv3 semantic future prediction and history retrieval with VAE dynamics, improving zero-shot LIBERO-Plus by 21.3 points and real-world visual-shift success from 25.8% to 61.5% Resources: Project · Code.
- 2026.7 · ACID: Action Consistency via Inverse Dynamics for Planning with World Models - Uses inverse dynamics to score whether imagined futures are consistent with candidate actions, improving world-model planning and trajectory selection.
- 2026.6 · Learning Transferable Dynamics Priors from Action to World Modeling - Introduces A2World, whose action-conditioned pretraining transfers dynamics priors to learned simulators and video-action policies in simulation and real robots.
- 2026.6 · World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis - Unifies predictive world modeling, language reasoning, and action synthesis so each capability can inform the others during robot decision making.
- 2026.6 · τ_0-WM: A Unified Video-Action World Model for Robotic Manipulation - Jointly predicts videos and actions, evaluates candidate futures, and rectifies control through simulator feedback within one manipulation world model.
- 2026.3 · Fast-WAM: Do World Action Models Need Test-time Future Imagination? - Studies when explicit future imagination is necessary and distills world-action knowledge into faster policies for efficient robot deployment Resources: Project · Code.
- 2026.1 · PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation - Scales predictive modeling in 3D point space, improving geometry-aware future prediction and manipulation across diverse real-world scenes and object configurations Resources: Project · Code.
- 2025.6 · V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning - Learns predictive video representations at scale and transfers them to visual understanding, anticipation, and planning for physical tasks Resources: Project · Code.
Verification and Safety Evaluation
Success detection, temporal and specification-grounded verification, runtime policy steering, execution checking, and process-level evaluation of physical safety. Papers whose primary contribution is diagnosis or repair are listed separately under Failure Detection and Recovery.
- 2026.8 · MANIGUARD: A Benchmark and Data Suite for Specification-Grounded Safety Evaluation and Improvement of Robotic Manipulation - Evaluates manipulation trajectories against explicit safety specifications and provides annotated data for improving process-level physical safety during execution.
- 2026.7 · SafeRelBench: A Spatial-Relation-Aware Benchmark for Process-Level Safety in VLM-Driven Embodied Agents - Measures safety violations caused by changing spatial relations during execution, revealing failures that final-state success metrics overlook during long-horizon tasks.
- 2026.7 · EgoSafetyBench: A Diagnostic Egocentric Video Benchmark for Evaluating Embodied VLMs as Runtime Safety Guards - Tests embodied VLMs as streaming runtime guards using temporally annotated egocentric scenarios and diagnostic categories of physical risk.
- 2026.6 · Visual Verification Enables Inference-time Steering and Autonomous Policy Improvement - Uses visual outcome verification to steer policy execution at inference time and generate signals for autonomous policy improvement.
- 2026.5 · Pre-VLA: Preemptive Runtime Verification for Reliable Vision-Language-Action and World-Model Rollouts - Verifies proposed VLA and world-model rollouts before execution, preventing unsafe or inconsistent actions rather than reacting after failure.
- 2026.4 · Open-Loop Planning, Closed-Loop Verification: Speculative Verification for VLA - Pairs efficient open-loop action planning with closed-loop speculative verification, preserving speed while catching execution drift and task errors.
- 2024.10 · AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation - Detects manipulation failures from visual observations and reasons about their causes, supporting interpretable diagnosis before recovery planning and replanning.
- 2023.3 · Vision-Language Models as Success Detectors - Evaluates general-purpose vision-language models as reusable task-success detectors, reducing the need for separately trained reward or verification models.
Failure Detection and Recovery
Failure prediction and detection, causal diagnosis, natural-language failure reasoning, motion correction, task-level recovery, backtracking, and safe human takeover.
- 2026.9 · Spotter: Let the Embodied Model Lead, and the VLM Reflect for It - Keeps a frozen VLA executing while a local screener monitors every action chunk in parallel; a VLM intervenes only on confirmed errors, repairs them with motion primitives, verifies the fix, then returns control. Resources: Project · Code.
- 2026.1 · CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding - Proactively detects uncertain subtasks, backtracks execution, and uses minimum Bayes risk decoding to select safer corrective VLA actions Resources: Project.
- 2025.10 · FailSafe-VLM: Reasoning and Recovery from Failures in Vision-Language-Action Models - Uses vision-language reasoning to identify VLA execution failures, explain their causes, and produce context-sensitive recovery strategies during long-horizon execution.
- 2025.9 · FPC-VLA: A Vision-Language-Action Framework with a Supervisor for Failure Prediction and Correction - Adds a supervisory model that predicts impending failures and corrects VLA actions before errors propagate through long-horizon tasks Resources: Project.
- 2025.6 · SAFE: Multitask Failure Detection for Vision-Language-Action Models - Trains a shared failure detector across manipulation tasks, improving recognition of execution errors without task-specific monitoring models or reward functions Resources: Project.
- 2024.9 · Automating Robot Failure Recovery Using Vision-Language Models - Uses vision-language models to interpret execution failures and automatically select recovery actions, establishing a practical general-purpose recovery pipeline.
Skill Learning and Tool Use
Robot skill repositories, physical and software tool use, capability discovery, task decomposition, skill composition, code generation, and execution interfaces such as ROS.
- 2026.9 · Agent Priors-guided Policy Learning - Uses each skill policy's training-time structural prior as its composition interface; a construction agent learns and verifies prior-specific policies from few demonstrations, and a runtime agent selects and composes them for new task goals. Resources: Project · Code.
- 2026.6 · CLASP: Language-Driven Robot Skill Selection and Composition using Task-Parameterized Learning - Uses language to select and compose task-parameterized robot skills, supporting adaptable execution beyond fixed primitive sequences and object configurations.
- 2026.3 · CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation - Benchmarks code-generating agents on robot manipulation and provides a framework for improving executable programs through environment feedback and iterative repair.
- 2025.5 · Dynamic Robot Tool Use with Vision Language Models - Uses vision-language reasoning to select, localize, and operate previously unseen physical tools under changing task requirements and scene conditions.
- 2025.5 · DeCo: Task Decomposition and Skill Composition for Zero-Shot Generalization in Long-Horizon 3D Manipulation - Decomposes language goals and composes learned 3D manipulation skills, enabling zero-shot generalization to new long-horizon task combinations and environments Resources: Project.
- 2025.5 · RAI: Flexible Agent Framework for Embodied AI - Provides a modular agent framework that connects language reasoning with ROS tools, robot skills, logging, and execution feedback Resources: Code.
- 2024.6 · ROS-LLM: A ROS framework for embodied AI with task feedback and structured reasoning - Integrates structured language-model reasoning with ROS execution and task feedback, exposing robot capabilities through a practical agent interface.
- 2022.9 · Code as Policies: Language Model Programs for Embodied Control - Generates executable policy code that composes perception and control APIs, establishing a foundational approach to language-driven robot tool use.
Long-Horizon Manipulation
Multi-stage manipulation systems that sustain task progress through hierarchical policies, memory, progress estimation, skill chaining, adaptive planning, reflection, and recovery-aware execution.
- 2026.6 · SERF: Spatiotemporal Environment and Robot Feature Map for Long-Horizon Mobile Manipulation - Maintains spatiotemporal environment and robot features to support persistent state tracking across long-horizon mobile manipulation tasks under partial observability.
- 2026.4 · Long-Horizon Manipulation via Trace-Conditioned VLA Planning - Conditions VLA planning on structured execution traces, preserving progress information and improving action selection over extended manipulation sequences.
- 2026.4 · Goal2Skill: Long-Horizon Manipulation with Adaptive Planning and Reflection - Converts goals into reusable skills and adapts their composition through execution feedback and reflection during long-horizon manipulation and unexpected state changes.
- 2026.3 · ProgressVLA: Progress-Guided Diffusion Policy for Vision-Language Robotic Manipulation - Models task progress explicitly and uses it to guide diffusion-based action generation across multi-stage language-conditioned manipulation and recovery.
- 2026.2 · LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies - Links object-centric policies into compositional plans, enabling VLA execution to reuse local behaviors across longer manipulation tasks and environments.
Embodied Navigation
Vision-language and object-goal navigation agents that combine spatial memory, map or scene reasoning, active exploration, adaptive control, self-correction, and cross-embodiment deployment.
- 2026.8 · Embodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation - Introduces TAMP-Nav, combining pixel-to-3D actions, selective reasoning, trajectory memory, and process-level alignment for efficient embodied navigation with lower overhead.
- 2026.6 · SpaceVLN: A Zero-Shot Vision-and-Language Navigation Agent with Online Spatial Cognitive Memory and Reasoning - Maintains online spatial cognitive memory and reasons over it to perform zero-shot language-guided navigation in unseen environments without task-specific training.
- 2026.3 · AgentVLN: Towards Agentic Vision-and-Language Navigation - Reframes vision-language navigation as an agentic loop with explicit reasoning, state tracking, and adaptive decisions during execution and failures.
- 2026.3 · SysNav: Multi-Level Systematic Cooperation Enables Real-World, Cross-Embodiment Object Navigation - Coordinates reasoning and control at multiple levels to support real-world object navigation across different robot embodiments and environment types.
- [2024.12 · [RSS 2025] Uni-NaVid: A Video-based Vision-Language-Action Model for Unifying Embodied Navigation Tasks](https://arxiv.org/abs/2412.06224) - Uses video-conditioned action prediction to unify multiple embodied navigation settings within a single vision-language-action model and deployment environments Resources: Project.
- 2022.7 · LM-Nav: Robotic Navigation with Large Pre-Trained Models of Language, Vision, and Action - Composes pretrained language, vision, and navigation models without task-specific fine-tuning, demonstrating modular zero-shot robot navigation across unseen environments.
- [2018.6 · [CVPR] Vision-and-Language Navigation: Interpreting Visually-Grounded Navigation Instructions in Real Environments](https://openaccess.thecvf.com/content_cvpr_2018/papers/Anderson_Vision-and-Language_Navigation_Interpreting_CVPR_2018_paper.pdf) - Introduces the Room-to-Room task and benchmark, establishing the standard formulation for grounding natural-language instructions in embodied navigation and evaluation.
Human-Robot Interaction
Language-enabled human feedback, dialogue, mixed initiative, human modeling, social robot planning, collaborative manipulation, and human-in-the-loop coordination.
- 2024.1 · Understanding LLM-powered Human-Robot Interaction - Examines how large language models reshape human-robot communication, task interpretation, trust, and interaction design in embodied settings and deployments.
- 2023.3 · Large Language Models as Zero-Shot Human Models for Human-Robot Interaction - Uses language models to predict human preferences and behavior without task-specific training, supporting adaptive planning in human-robot interaction.
Governance and Physical Safety
Governance constraints, policy-compliant execution, consent and delegation, runtime guardrails, physical irreversibility, and mechanisms for keeping autonomous behavior within acceptable boundaries.
- 2026.5 · Consent Chain Degradation in Embodied Multi-Agent Systems - Studies how consent constraints weaken across delegated multi-agent execution, exposing governance risks unique to embodied action chains and delegation structures.
- 2026.4 · EmbodiedGovBench: A Benchmark for Governance-Constrained Embodied Agents - Evaluates whether embodied agents follow governance policies while pursuing tasks, separating rule compliance from conventional completion metrics or isolated benchmarks.
- 2026.4 · Harnessing Embodied Agents: Runtime Governance for Policy-Constrained Execution - Introduces runtime governance mechanisms that monitor and constrain agent actions against explicit policies during physical execution with minimal latency.
Benchmarks and Datasets
Benchmarks, simulators, and datasets for generalist policies, long-horizon tasks, lifelong transfer, navigation, multi-robot systems, manipulation, recovery, and embodied safety.
2026.10 · RoboQuest: Generalist Physical Agents that Search, Inspect and Test - Evaluates frontier multimodal agents that drive a simulated mobile manipulator through tool calls on ten kitchen tasks requiring searching, inspecting, or testing to uncover hidden information before submitting. Resources: Project · Code · Dataset.
2026.9 · EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks - Evaluates how agents retain visual details, track changing world states, remember interaction outcomes, and reuse experience through 2,554 executable episodes, with an external memory-system baseline. Resources: Project · Code · Dataset.
2026.3 · RoboCasa365: A Large-Scale Simulation Framework for Generalist Robot Policies - Expands everyday manipulation simulation with broad tasks and environments for training and evaluating generalist, long-horizon robot policies across varied settings.
2025.4 · RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins - Uses generative digital twins to scale diverse dual-arm manipulation tasks, data collection, and reproducible policy evaluation across varied scenes.
2023.10 · Habitat 3.0: A Co-Habitat for Humans, Avatars and Robots - Provides interactive human-avatar-robot simulation for studying navigation, collaboration, and socially situated embodied agents at scale across collaborative tasks.
2023.10 · Open X-Embodiment: Robotic Learning Datasets and RT-X Models - Aggregates cross-institution robot datasets and RT-X models, establishing a shared foundation for cross-embodiment learning, transfer, and generalization at scale Resources: Project / Dataset · Code.
2023.6 · LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning - Evaluates lifelong robot learning across task suites designed to test knowledge transfer, retention, and efficient adaptation across sequential tasks.
2021.12 · CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks - Benchmarks language-conditioned policies on diverse, compositional manipulation chains, making long-horizon completion and generalization directly measurable across extended task sequences.
Foundations and Industrial Systems
Openly released robot foundations and industrial technical systems that materially shape reusable generalist control, embodied reasoning, cross-embodiment transfer, or deployment practice.
- [2025.3 · [Google DeepMind] Gemini Robotics: Bringing AI into the Physical World](https://arxiv.org/abs/2503.20020) - Presents an industrial-scale embodied reasoning model and generalist VLA system designed to transfer semantic understanding into dexterous physical action.
- 2024.5 · Octo: An Open-Source Generalist Robot Policy - Provides an openly released generalist robot policy, training stack, and checkpoints built from heterogeneous manipulation datasets across embodiments Resources: Project · Code.
Extended Reading
Papers moved out of the main list during the 2026-08-27 curation pass remain available in extended-reading.md. They are useful adjacent work, but scored lower on direct agentic relevance, influence, evidence, representativeness, or resource completeness.
Related Awesome Lists
Citation
If you find this repository useful, please consider citing this list:
@misc{liu2026awesomeagenticrobotics,
title = {Awesome Agentic Robotics},
author = {Duo Liu and Xinbai Wang},
journal = {GitHub repository},
url = {https://github.com/w-xb/awesome-agentic-robotics},
year = {2026},
}
Contributing
Contributions are welcome. Please read contributing.md before submitting a pull request or using the paper-suggestion issue template.