Skip to content
Projects

w-xb/awesome-agentic-robotics

View on GitHub

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.

How these counts are measured

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

Learning

README.md

Original

Awesome Agentic Robotics Awesome

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

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.

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.

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.

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.

Embodied Memory

Memory systems for physical agents, including working and episodic memory, semantic-spatial scene representations, experience retrieval, and persistent world or task state.

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.

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.

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.

Failure Detection and Recovery

Failure prediction and detection, causal diagnosis, natural-language failure reasoning, motion correction, task-level recovery, backtracking, and safe human takeover.

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.

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.

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.

Human-Robot Interaction

Language-enabled human feedback, dialogue, mixed initiative, human modeling, social robot planning, collaborative manipulation, and human-in-the-loop coordination.

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.

Benchmarks and Datasets

Benchmarks, simulators, and datasets for generalist policies, long-horizon tasks, lifelong transfer, navigation, multi-robot systems, manipulation, recovery, and embodied safety.

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.