Awesome Agentic Robotics

This is the companion repository for the survey A Survey of Agentic Robotics: Toward Continual Self-Improvement, by Jiaming Wang, Yuhua Jiang, Zhengcheng Shen, and Harold Soh.

Paper: alphaXiv. Cite: see Citation.

In classical robotics, engineers decompose tasks, specify objectives, write control code, and repair failures before deployment. In agentic robotics, a foundation model (FM), usually a large language model (LLM) or vision-language model (VLM), makes some of these decisions at run time or while the system improves. This repository lists the systems analyzed in the survey and organizes them with the survey's taxonomy: where the FM takes part in generating behavior, which improvement decisions it makes, and how feedback is retained.

The architecture of agentic robotics used in the survey. In the agent–environment loop, task orchestration decides what to do and action generation decides how to do it; verification and memory close the loop. A slower self-improvement loop uses accumulated experience to change persistent components, which changes future behavior. Individual systems implement only some of these modules.

Contents

• Taxonomy
• Key findings
• Paper list
  • Task orchestration (T)
  • Motion/action generation (M)
  • Task and motion (T + M)
  • Fixed-procedure improvement (I_fixed)
  • Agent-directed improvement (I_agent)
• Benchmarks
• Open challenges
• Related surveys
• Survey data
• Citation
• Contributing

Taxonomy

Three loci of FM participation

The survey codes each system by what the FM actually outputs or decides in the system as evaluated, not by how the system is described. One FM may act at several loci, and a system need not use all of them.

Locus	The FM decides...	Coded when the FM output contains...	Representative systems
T: task orchestration	what behavior to pursue next	a subgoal, plan, skill/tool/API call, or semantic target (an object or place) that other modules turn into geometry	SayCan, ProgPrompt, SayPlan, KnowNo, COME-Robot
M: motion/action generation	how the behavior is physically realized	actions or a chosen action candidate, poses, waypoints, keypoints, constraints, cost maps, rewards optimized at execution time, metric skill parameters, or code that computes them; or the actions of an FM policy	PIVOT, VoxPoser, ReKep, Code as Policies, RoboCat
Ifixed: fixed-procedure improvement	the content of a candidate inside a prescribed update	a new policy, reward, program, or skill (or the data for it), while the designer fixes which artifact changes, what evidence the FM sees, how candidates are tested, and the update procedure	Scaling Up & Distilling Down, RoboCat, Eureka, CaP-X, Act-Observe-Rewrite
Iagent: agent-directed improvement	also how the improvement proceeds	a candidate plus at least one of the process decisions D1–D4 below	RHO, ASPIRE, ENPIRE

Who makes the improvement decisions?

Both kinds of improvement may use an FM to write the new policy, program, reward, or skill. They differ in five observable decisions:

• D1 which kind of artifact to change (policy, reward, training procedure, or data);
• D2 what evidence to collect;
• D3 which tests or experiments to run;
• D4 how the update procedure works;
• D5 whether a candidate is kept.

A system is I_agent if the FM makes at least one of D1–D4. D5 does not enter the label, because a fixed acceptance rule is often a deliberate safety choice.

Feedback-closure regimes

How verified feedback influences later behavior:

Regime	What is kept
Open loop	No execution feedback is used.
Step-level	Feedback shapes only the next decision.
Within-rollout memory	State, execution traces, or diagnoses are kept for the rest of the current task instance.
Cross-rollout memory	Records (lessons, corrections, trajectories) are retrieved or replayed in later task instances.
Persistent modification	An executable or learned artifact (weights, reward, program, skill, planning model) is changed and used in place of the old one.

The regime describes how feedback is retained. It does not rank systems: a step-level controller can be more reliable than a system that learns across rollouts.

Evolution of FM participation

Representative systems by year of first public version, marked by the loci at which FMs influence robot behavior (coded from full text). Outlined I: fixed-procedure improvement; filled I: agent-directed improvement. The figure shows a trend, not a ranking.

Key findings

The survey ran a systematic search of arXiv and OpenAlex (January 2022 to September 2026), screened the results, and coded all 2,343 systems within its scope. It then read the full texts of 78 of them in depth.

• FM decisions are moving from tasks to motion. In 2022–2023, FMs mostly planned or selected skills above a fixed execution stack. Later systems also generate programs, geometric objectives, keypoints, constraints, and actions. From full-text coding, the share of systems in which the FM produces motion-level outputs rose from 27% in 2023 to 63% in 2026. Task-level participation stayed near three quarters.
• Persistent improvement is not new, but agent-directed improvement is. In every year, about one system in five improves persistently, mostly through fixed data-collection, distillation, reward-search, or learning pipelines. Only 17 of the 2,343 systems let the FM direct the improvement, and 16 of them appeared in 2026.
• The delegated decisions come separately. Among the 17 agent-directed systems, the FM chooses its evidence in 15, its tests in 10, the update procedure in 7, and the kind of artifact in 7.
• Persistent modification is not yet continual improvement. Representative self-improving systems run a single improvement iteration (RoboCat) or produce one artifact per task (Eureka, CaP-X, Act-Observe-Rewrite, ASPIRE). Among the 17 agent-directed systems, only one re-runs earlier regression tests before it promotes a change, 7 test on held-out tasks, and 8 report real-robot results.
• There is no evidence yet that more FM authority pays off. The only equal-budget comparison with fixed-procedure loops (HARBOR against Eureka and REvolve) ran in simulation and gave the agent human heuristics that the baselines lacked.
• One role label hides different systems. Maestro, CaP-X, Act-Observe-Rewrite, and RHO are all "coding agents". However, they differ in whether the FM is called on every trial, whether the agent tests its own edits, and who decides that a change is kept.

Share of the 2,343-system corpus at each locus by year, coded from abstracts. Abstracts understate motion-level outputs. Diamonds show the motion-level share corrected with full-text coding of a random sample. 2026 covers January to September.

Paper list

Each table is sorted newest first. Year is the year of the first public version (usually arXiv v1). Evidence says whether the paper reports simulated experiments, physical-robot experiments, or both.

• Analyzed in the survey: the 78 systems whose full texts were coded for the survey (the per-system table in the paper's appendix). For each one, the table gives its feedback-closure regime and its main feedback-closure mechanism. Systems that also improve persistently are listed under the two improvement sections.
• More systems: other papers within the survey's scope, placed by the abstract-level coding of the survey corpus.

Task orchestration (T)

The FM decides what to do next: subgoals, plans, skill or API calls, or semantic targets. Other modules turn these decisions into motion. Systems make task-level decisions executable in four ways: they restrict the FM to exposed skills, ground its choices in the current state, rely on a downstream solver, or recover from execution feedback.

Analyzed in the survey (27)

Year	System and paper	Venue	Evidence	Feedback regime	Feedback closure
2026	ACE — Agentic Control for Embodied Manipulation via Zero-Shot Workflow Reasoning	arXiv	Physical	Within-rollout memory	Verification, repair, and replanning of sub-goals.
2026	HoloMind / LongAct — When Robots Do the Chores: A Benchmark and Agent for Long-Horizon Household Task Execution	arXiv	Simulation	Cross-rollout memory	Episodic experience memory.
2024	COME-Robot — Closed-Loop Open-Vocabulary Mobile Manipulation with GPT-4V	ICRA 2025	Physical	Within-rollout memory	Verification and hierarchical recovery.
2024	ORGANA — A Robotic Assistant for Automated Chemistry Experimentation and Characterization	Matter 2025	Physical	Within-rollout memory	LLM plans; records summarized within one campaign.
2024	ChemAgents — A Multiagent-Driven Robotic AI Chemist Enabling Autonomous Chemical Research on Demand	JACS 2025	Physical	Within-rollout memory	Multi-agent LLM compiles lab-robot workflows.
2023	CLAIRify — Large Language Models for Chemistry Robotics	Autonomous Robots 2023	Physical	Step-level	Verified LLM task plans for a chemistry lab robot.
2023	SMART-LLM — Smart Multi-Agent Robot Task Planning Using Large Language Models	IROS 2024	Sim + physical	Open loop	LLM multi-robot decomposition and allocation.
2023	PG-VLM — Physically Grounded Vision-Language Models for Robotic Manipulation	ICRA 2024	Physical	Open loop	Physically grounded VLM answers planner queries.
2023	ConceptGraphs — Open-Vocabulary 3D Scene Graphs for Perception and Planning	ICRA 2024	Sim + physical	Step-level	LLM selects objects from 3D scene-graph captions.
2023	SayPlan — Grounding Large Language Models Using 3D Scene Graphs for Scalable Robot Task Planning	CoRL 2023	Simulation	Within-rollout memory	Scene-graph simulator feedback and iterative replanning.
2023	KnowNo — Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners	CoRL 2023	Sim + physical	Step-level	Uncertainty-triggered help seeking.
2023	DoReMi — Grounding Language Model by Detecting and Recovering from Plan-Execution Misalignment	IROS 2024	Sim + physical	Within-rollout memory	Constraint-violation detection and recovery.
2023	Statler — State-Maintaining Language Models for Embodied Reasoning	ICRA 2024	Sim + physical	Within-rollout memory	Maintained world-state memory.
2023	REFLECT — Summarizing Robot Experiences for Failure Explanation and Correction	CoRL 2023	Sim + physical	Within-rollout memory	Failure diagnosis and correction planning.
2023	AutoTAMP — Autoregressive Task and Motion Planning with LLMs as Translators and Checkers	ICRA 2024	Sim + physical	Open loop	LLM translates language into STL for TAMP.
2023	TidyBot — Personalized Robot Assistance with Large Language Models	Autonomous Robots 2023	Physical	Step-level	LLM-summarized user preferences for object placement.
2023	L3MVN — Leveraging Large Language Models for Visual Target Navigation	IROS 2023	Sim + physical	Step-level	LM scores each frontier; argmax picks the subgoal.
2023	Text2Motion — From Natural Language Instructions to Feasible Plans	Autonomous Robots 2023	Simulation	Step-level	LLM skill plans; learned feasibility and STAP compute geometry.
2023	PaLM-E — An Embodied Multimodal Language Model	ICML 2023	Sim + physical	Step-level	Re-prompted each step with image and past steps.
2022	LLM-Planner — Few-Shot Grounded Planning for Embodied Agents with Large Language Models	ICCV 2023	Simulation	Within-rollout memory	Few-shot LLM subgoal plans, re-planned with observed objects.
2022	HULC++ — Grounding Language with Visual Affordances over Unstructured Data	ICRA 2023	Physical	Open loop	GPT-3 subgoal sequences for a learned policy.
2022	ProgPrompt — Generating Situated Robot Task Plans Using Large Language Models	ICRA 2023	Sim + physical	Step-level	Assertion checks with scripted recovery branches.
2022	NLMap — Open-vocabulary Queryable Scene Representations for Real World Planning	ICRA 2023	Physical	Step-level	LLM queries an open-vocabulary scene representation.
2022	LM-Nav — Robotic Navigation with Large Pre-Trained Models of Language, Vision, and Action	CoRL 2022	Physical	Open loop	LLM landmark extraction grounded by CLIP.
2022	Inner Monologue — Embodied Reasoning through Planning with Language Models	CoRL 2022	Sim + physical	Within-rollout memory	Success, scene, and human feedback in the planning prompt.
2022	SayCan — Do As I Can, Not As I Say: Grounding Language in Robotic Affordances	CoRL 2022	Physical	Step-level	Affordance-weighted skill selection from the current state.
2022	Zero-Shot Planner — Language Models as Zero-Shot Planners: Extracting Actionable Knowledge for Embodied Agents	ICML 2022	Simulation	Open loop	Plan generation mapped to admissible actions.

More systems

Year	System and paper	Venue	Evidence	Summary
2026	Closed-Loop Multi-Agent Manipulation — A Closed-Loop Multi-Agent Framework for Robust Multi-Robot Manipulation	RSS 2026	Physical	A planning agent, per-robot manipulation agents, and a verification agent coordinate tool use and feed semantic corrections back after physical execution.
2026	Language-to-Action Evaluation — From Language to Action: Can LLM-Based Agents Be Used for Embodied Robot Cognition?	arXiv	Simulation	The study evaluates an LLM cognitive core with working and episodic memory and exposes persistent grounding and reliability limitations.
2026	Agentic AI for Robot Control — Flexible but still Fragile	arXiv	Physical	A physical-robot study stress-tests iterative tool-based planners and documents how operator intervention, timing, and execution events still cause failures.
2025	ASP — Agentic Scene Policies: Unifying Space, Semantics, and Affordances for Robot Action	arXiv	Physical	ASP lets an LLM agent query an object-centric scene representation for semantic, spatial, and affordance facts that guide downstream motion planning.
2025	L3M+P — Lifelong Planning with Large Language Models	arXiv	Sim + physical	L3M+P maintains a verified knowledge graph from sensor and language updates and retrieves it to construct classical planning problems.
2025	Memory-Augmented Household Agent — LLM-Empowered Embodied Agent for Memory-Augmented Task Planning in Household Robotics	Austrian Robotics Workshop 2025	Physical	Routing, planning, and knowledge agents use retrieval-augmented long-term object memory to plan household service-robot tasks.
2025	Reflective Planning — Vision-Language Models for Multi-Stage Long-Horizon Robotic Manipulation	CoRL 2025	Simulation	A VLM uses a learned dynamics model to imagine future states and reflect on suboptimal actions before executing multi-stage manipulation plans.
2024	ROSA — Robot Operating System Agent	IEEE Aerospace 2025	Sim + physical	ROSA exposes validated ROS 1 and ROS 2 operations as agent tools and adds constraint checks for natural-language robot inspection and control.
2024	SELP — Generating Safe and Efficient Task Plans for Robot Agents with Large Language Models	arXiv	Simulation	SELP maps language to temporal logic using equivalence voting and constrains decoding so generated drone and manipulation plans satisfy user rules.
2024	MHRC — Closed-loop Decentralized Multi-Heterogeneous Robot Collaboration with Large Language Models	arXiv	Simulation	MHRC lets heterogeneous mobile and manipulation agents exchange textual feedback, request help, and revise decentralized collaborative plans.
2024	AutoRT — Embodied Foundation Models for Large Scale Orchestration of Robotic Agents	arXiv	Physical	AutoRT combines VLM scene understanding, LLM task proposals, and safety filters to orchestrate a robot fleet that collected 77,000 real episodes.
2023	Instruct2Act — Mapping Multi-modality Instructions to Robotic Actions with Large Language Model	arXiv	Physical	Instruct2Act generates Python programs that call foundation-model perception APIs and robot primitives directly from multimodal instructions.
2023	Demo2Code — From Summarizing Demonstrations to Synthesizing Code via Extended Chain-of-Thought	NeurIPS 2023	Simulation	Demo2Code recursively summarizes demonstrations into reusable behavior specifications and expands them into executable robot programs.
2023	Grounded Decoding — Guiding Text Generation with Grounded Models for Robot Control	NeurIPS 2023	Sim + physical	Grounded Decoding combines token probabilities with affordance, safety, and preference model scores while an LLM generates a robot plan.

Motion/action generation (M)

The FM outputs information that directly governs physical realization. For example, it may select an action candidate, specify a cost or reward for a solver, or choose the next viewpoint.

Analyzed in the survey (6)

Year	System and paper	Venue	Evidence	Feedback regime	Feedback closure
2025	SAS-Prompt — Large Language Models as Numerical Optimizers for Robot Self-Improvement	ICRA 2025	Sim + physical	Cross-rollout memory	In-context controller-parameter search from rollout traces.
2024	VLMnav — End-to-End Navigation with Vision-Language Models: Transforming Spatial Reasoning into Question-Answering	NeSy 2025	Simulation	Step-level	Step-wise visual navigation decisions.
2024	VLM-Social-Nav — Socially Aware Robot Navigation Through Scoring Using Vision-Language Models	IEEE RA-L 2025	Physical	Step-level	VLM-derived social cost in per-step replanning.
2024	PIVOT — Iterative Visual Prompting Elicits Actionable Knowledge for VLMs	ICML 2024	Sim + physical	Step-level	Iterative visual action-candidate selection.
2023	Language to Rewards — Language to Rewards for Robotic Skill Synthesis	CoRL 2023	Sim + physical	Step-level	Reward-mediated skill realization via MPC.
2023	NavGPT — Explicit Reasoning in Vision-and-Language Navigation with Large Language Models	AAAI 2024	Simulation	Within-rollout memory	LLM picks the next viewpoint from described views and history.

Task and motion (T + M)

The FM chooses the behavior and also specifies how it is realized. It does this through executable programs, geometric constraints, value maps, or parameterized skill calls.

Analyzed in the survey (21)

Year	System and paper	Venue	Evidence	Feedback regime	Feedback closure
2026	Uni-Skill — Building Self-Evolving Skill Repository for Generalizable Robotic Manipulation	ICRA 2026	Sim + physical	Open loop	On-demand skill retrieval from an offline skill repository.
2026	ALRM — Agentic LLM for Robotic Manipulation	arXiv	Simulation	Within-rollout memory	ReAct-style tool feedback.
2025	Maestro — Orchestrating Robotics Modules with Vision-Language Models for Zero-Shot Generalist Robots	NeurIPS SpaVLE Workshop 2025	Physical	Cross-rollout memory	Cross-trial in-context reflection.
2025	Reliable Code-as-Policies — Towards Reliable Code-as-Policies: A Neuro-Symbolic Framework for Embodied Task Planning	NeurIPS 2025	Sim + physical	Within-rollout memory	Interactive validation during code generation.
2025	ManiAgent — An Agentic Framework for General Robotic Manipulation	arXiv	Sim + physical	Within-rollout memory	Cross-task action-sequence cache (retrieval memory).
2025	PragmaBot — A Pragmatist Robot: Learning to Plan Tasks by Experiencing the Real World	IEEE RA-L 2026	Physical	Cross-rollout memory	Experiential memory summarized from trial and error.
2025	ARNA — General-Purpose Robotic Navigation via LVLM-Orchestrated Perception, Reasoning, and Acting	arXiv	Simulation	Within-rollout memory	Iterative tool-use feedback.
2025	RAI — Flexible Agent Framework for Embodied AI	PAAMS 2025	Sim + physical	Within-rollout memory	Robot-tool feedback, including point targets.
2025	Agentic Robot — A Brain-Inspired Framework for Vision-Language-Action Models in Embodied Agents	arXiv	Simulation	Within-rollout memory	Verifier triggers a fixed recovery routine and retry.
2025	ELLMER — Embodied Large Language Models Enable Robots to Complete Complex Tasks in Unpredictable Environments	Nature Machine Intelligence 2025	Physical	Step-level	GPT-4 writes retrieved code with force/vision feedback.
2024	BUMBLE — Unifying Reasoning and Acting with Vision-Language Models for Building-Wide Mobile Manipulation	ICRA 2025	Physical	Cross-rollout memory	Short- and long-term episodic memory.
2024	ReKep — Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation	CoRL 2024	Sim + physical	Step-level	Re-optimization of relational keypoint constraints.
2023	DROC — Distilling and Retrieving Generalizable Knowledge for Robot Manipulation via Language Corrections	ICRA 2024	Physical	Cross-rollout memory	Retrieval of knowledge distilled from language corrections.
2023	ChatGPT Research Group — ChatGPT Research Group for Optimizing the Crystallinity of MOFs and COFs	ACS Central Science 2023	Physical	Cross-rollout memory	GPT-4 writes a liquid-handling script; BO picks conditions.
2023	Incremental humanoid learning — Incremental Learning of Humanoid Robot Behavior from Natural Interaction and Large Language Models	Frontiers in Robotics and AI 2024	Sim + physical	Cross-rollout memory	Corrected interactions stored and retrieved.
2023	VoxPoser — Composable 3D Value Maps for Robotic Manipulation with Language Models	CoRL 2023	Sim + physical	Step-level	Geometric replanning over composed value maps.
2023	RoCo — Dialectic Multi-Robot Collaboration with Large Language Models	ICRA 2024	Sim + physical	Within-rollout memory	Dialog-based coordination and collision-aware replanning.
2023	RoboGPT (construction) — Robot-Enabled Construction Assembly with Automated Sequence Planning Based on ChatGPT: RoboGPT	Buildings 2023	Sim + physical	Open loop	ChatGPT assembly sequences with pipe coordinates.
2023	LLM-GROP — Task and Motion Planning with Large Language Models for Object Rearrangement	IROS 2023	Sim + physical	Open loop	LLM arrangement offsets instantiated by TAMP.
2022	VLMaps — Visual Language Maps for Robot Navigation	ICRA 2023	Sim + physical	Open loop	LLM writes navigation code with landmark goals and metric moves.
2022	Code as Policies — Language Model Programs for Embodied Control	ICRA 2023	Sim + physical	Step-level	Reactive perception–action loops in generated code.

More systems

Year	System and paper	Venue	Evidence	Summary
2026	ARCHITECT — A Few Words Go a Long Way: Language Guided Robot Policy Synthesis	arXiv	Physical	ARCHITECT treats policy acquisition as interactive program synthesis and distills trace-grounded language corrections into a persistent skill library.
2025	ModuLoop — Low-Level Code Generation using Modular Synthesizer and Closed-Loop Debugger for Robotic Control	IEEE RA-L 2025	Physical	ModuLoop generates low-level control code module by module, inserts diagnostic probes, and iteratively debugs executions for calibration and manipulation.
2024	MALMM — Multi-Agent Large Language Models for Zero-Shot Robotic Manipulation	IROS 2025	Sim + physical	MALMM separates high-level planning, low-level code generation, and supervision among specialist LLM agents that replan from each observation.
2024	IFVF Code Generation — Enabling Robots to Follow Abstract Instructions and Complete Complex Dynamic Tasks	arXiv	Physical	GPT-4 retrieves domain knowledge and writes robot code whose execution is adapted through integrated force and visual feedback.
2024	LABOR — Large Language Models for Orchestrating Bimanual Robots	arXiv	Simulation	LABOR uses an LLM to analyze task dependencies and synthesize coordination policies for two robot arms without bimanual demonstrations.
2024	RobotScript — An LLM and Simulator-Assisted Robot Programming System for Human-Robot Interaction	arXiv	Sim + physical	RobotScript maps free-form instructions to a deployable robot API, validates generated programs in simulation, and transfers them to physical arms.
2023	ChatGPT for Robotics — Design Principles and Model Abilities	IEEE Access 2024	Sim + physical	The framework exposes a high-level robot function library and uses dialogue, code generation, and closed-loop feedback to refine behavior across embodiments.

Fixed-procedure improvement (I_fixed)

The system changes a persistent artifact, such as policy weights, a reward program, a skill library, or a planning model. The FM supplies candidates inside an update procedure that the designer fixes.

Analyzed in the survey (21)

Year	System and paper	Venue	Evidence	Loci	Feedback closure
2026	CaP-X / CaP-Agent0 — A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation	ICML 2026	Sim + physical	T, M, Ifixed	One-pass skill-library synthesis from successful rollouts.
2026	Act-Observe-Rewrite — Multimodal Coding Agents as In-Context Policy Learners for Robot Manipulation	arXiv	Simulation	T, M, Ifixed	Policy-code rewriting from a fixed outcome package.
2025	Self-Improving EFM — Self-Improving Embodied Foundation Models	NeurIPS 2025	Sim + physical	M, Ifixed	Online policy update from self-predicted rewards.
2025	VLMgineer — Vision-Language Models as Robotic Toolsmiths	ICLR 2026	Sim + physical	M, Ifixed	Evolutionary tool/action co-design.
2025	LAMARL — LLM-Aided Multi-Agent Reinforcement Learning for Cooperative Policy Generation	IEEE RA-L 2025	Sim + physical	M, Ifixed	LLM prior policy and reward code for MARL.
2024	LLM locomotion design — Leveraging Large Language Models for Comprehensive Locomotion Control in Humanoid Robots Design	Biomimetic Intelligence and Robotics 2024	Simulation	M, Ifixed	LLM designs trajectories, IK, and reward code.
2024	LASP — Language-Augmented Symbolic Planner for Open-World Task Planning	RSS 2024	Simulation	T, Ifixed	LLM diagnoses errors; fixed loop edits the planning domain.
2024	DrEureka — Language Model Guided Sim-to-Real Transfer	RSS 2024	Sim-to-real	Ifixed	LLM reward search; randomization ranges from a prior.
2024	AIC MLLM — Autonomous Interactive Correction MLLM for Robust Robotic Manipulation	CoRL 2024	Sim + physical	M, Ifixed	Pose correction and test-time model adaptation.
2024	InterPreT — Interactive Predicate Learning from Language Feedback for Generalizable Task Planning	RSS 2024	Sim + physical	T, Ifixed	Predicates from language feedback compiled into PDDL.
2024	Agentic Skill Discovery	Robotics and Autonomous Systems 2026	Simulation	T, Ifixed	Skill-library expansion under fixed RL and reward evolution.
2024	MOKA — Open-World Robotic Manipulation through Mark-Based Visual Prompting	RSS 2024	Physical	T, M, Ifixed	Distillation of collected experience into a policy.
2023	RoboGen — Towards Unleashing Infinite Data for Automated Robot Learning via Generative Simulation	ICML 2024	Simulation	T, Ifixed	Skill acquisition under a prescribed learning pipeline.
2023	UniSim — Learning Interactive Real-World Simulators	ICLR 2024	Sim + physical	T, M, Ifixed	RL fine-tuning of a VLA policy in a learned video simulator.
2023	Gen2Sim — Scaling up Robot Learning in Simulation with Generative Models	ICRA 2024	Simulation	T, Ifixed	LLM task decomposition and reward code for RL.
2023	Eureka — Human-Level Reward Design via Coding Large Language Models	ICLR 2024	Simulation	Ifixed	Evolutionary reward-program search.
2023	HALP — Lifelong Robot Learning with Human Assisted Language Planners	ICRA 2024	Sim + physical	T, Ifixed	LLM planner requests and reuses newly taught skills.
2023	Scaling Up & Distilling Down — Language-Guided Robot Skill Acquisition	CoRL 2023	Sim + physical	T, Ifixed	Policy distillation from LLM-guided data collection.
2023	RoboCat — A Self-Improving Generalist Agent for Robotic Manipulation	TMLR 2024	Sim + physical	M, Ifixed	Self-generated data for the next policy generation.
2023	LLM-Exo — LLM-Enabled Incremental Learning Framework for Hand Exoskeleton Control	IEEE T-ASE 2025	Physical	T, M, Ifixed	LLM writes new gesture code into exoskeleton command tables.
2022	ELM — Evolution Through Large Models	Handbook of Evolutionary Machine Learning 2024	Simulation	M, Ifixed	LLM mutation operator in MAP-Elites robot evolution.

More systems

Year	System and paper	Venue	Evidence	Summary
2025	Growing with Your Embodied Agent — A Human-in-the-Loop Lifelong Code Generation Framework for Long-Horizon Manipulation Skills	arXiv	Sim + physical	The framework converts human corrections into reusable code skills stored in external memory and retrieves them with task-specific hints.
2025	UROSA — Distributed AI Agents for Cognitive Underwater Robot Autonomy	arXiv	Sim + physical	UROSA distributes perception, reasoning, planning, adaptation, and on-the-fly ROS 2 node generation across specialized underwater-robot agents.
2024	RoboCodeX — Multimodal Code Generation for Robotic Behavior Synthesis	arXiv	Sim + physical	RoboCodeX uses tree-structured multimodal code generation to decompose instructions into object-centric, affordance-aware, and safety-aware programs.
2023	GenSim — Generating Robotic Simulation Tasks via Large Language Models	ICLR 2024	Sim-to-real	GenSim uses goal-directed and exploratory LLM agents to generate task assets, environment code, and demonstrations that train transferable policies.

Agent-directed improvement (I_agent)

The FM also decides how the improvement proceeds. This table lists all 17 agent-directed systems in the survey corpus, with the improvement decisions the FM makes and how each system was evaluated. † marks the systems that are also analyzed in depth in the survey.

● the FM decides; ○ fixed by the designer; – only one artifact can change. D1 kind of artifact to change; D2 evidence to collect; D3 tests or experiments to run; D4 update procedure; D5 who accepts a candidate (NR: not reported). Held-out test: evaluation on tasks, configurations, or environments not used during improvement. Regression test: re-testing earlier tasks or capabilities, either before a change is promoted or only after the update.

Year	System and paper	Venue	Evidence	D1	D2	D3	D4	Acceptance (D5)	Held-out test	Regression test
2026	WetRobo — A Reproducible Robot Kit for Coding Agents in Biological Laboratories	arXiv	Physical	●	●	●	●	FM	No	No
2026	Zetta ζ — An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence	arXiv	Simulation	–	●	○	○	Fixed rule	Yes	After update
2026	You Don't Need To Stay in The Loop: An Agentic Robotics Loop for Robot-Policy Improvement	arXiv	Simulation	●	●	●	●	Fixed rule	Yes	No
2026	Revisiting the "Push-T" Robot Manipulation Task with Agentic Robotics	arXiv	Simulation	–	●	●	●	FM	No	After update
2026	RoboHarness — Memory-Driven Orchestration of Heterogeneous Robot Policies for Long-Horizon Planning	arXiv	Sim + physical	●	○	○	○	Fixed rule	No	No
2026	GaP — A Graph-as-Policy Multi-Agent Self-Learning Harness For Variational Automation Tasks	arXiv	Sim + physical	–	●	○	○	Fixed rule	Yes	No
2026	ASPIRE — Agentic /Skills Discovery for Robotics †	arXiv	Sim + physical	–	●	●	○	Fixed rule	Yes	No
2026	RHO — Your Coding Agent is Secretly a Roboticist †	arXiv	Simulation	–	●	●	○	Fixed rule	Yes	After update
2026	Playful Agentic Robot Learning	arXiv	Sim + physical	–	●	○	○	Fixed rule	Yes	No
2026	HARBOR — A Harness Framework for Agentic Robot Reinforcement Learning	arXiv	Sim + physical	○	●	●	●	Fixed rule	No	No
2026	Heuristic Learning — From False Positives to Failure Topologies: Heuristic Learning for Auditable Dexterous Grasping Development	Zenodo	Simulation	●	●	●	●	Human	Yes	Before promotion
2026	ENPIRE — Agentic Robot Policy Self-Improvement in the Real World †	CoRL 2026	Sim + physical	●	●	●	●	FM	No	No
2026	Automating the Design of Embodied Agent Architectures	arXiv	Simulation	–	●	●	○	Fixed rule	No	No
2026	LITHE — Bridging Best-Effort Python and Real-Time C++ for Hot-Swapping Robotic Control Laws on Commodity Linux	arXiv	Physical	–	●	○	○	Fixed rule	No	No
2026	FATE — Closed-Loop Feasibility-Aware Task Generation with Active Repair for Physically Grounded Robotic Curricula	arXiv	Simulation	●	○	○	○	FM	No	No
2026	Agent-Driven Autonomous Reinforcement Learning Research: Iterative Policy Improvement for Quadruped Locomotion	arXiv	Simulation	●	●	●	●	FM	No	No
2025	Learning a High-quality Robotic Wiping Policy Using Systematic Reward Analysis and Visual-Language Model Based Curriculum	arXiv	Simulation	–	●	○	○	Fixed rule	No	No

Benchmarks

There is no shared benchmark for self-improvement yet. These are reusable benchmarks and evaluation substrates relevant to agentic robotics. "Feedback-rich" marks benchmarks that expose execution history, failures, or outcome evidence. "I substrate" means that an environment can support repeated adaptation but does not define a self-improvement protocol.

Benchmark	Locus	Environment	What it exposes
VirtualHome	T	Unity household simulator with symbolic actions	Executability and goal satisfaction of long-horizon action sequences; used by early LLM planners such as Zero-Shot Planner and ProgPrompt.
Embodied Agent Interface	T	VirtualHome + BEHAVIOR	Goal interpretation, subgoal decomposition, action sequencing, and transition modeling, with fine-grained planning and precondition errors.
LoHoRavens	T; feedback-rich	Language-conditioned tabletop simulation	Long-horizon reasoning over color, size, spatial relations, and arithmetic, and how execution observations are returned to an LLM planner.
LongAct	T; feedback-rich	Household tasks with low-level control abstracted	Instruction understanding, dependency management, persistent memory, and replanning; an improvement-rate metric measures gains from accumulated experience.
EmbodiedBench	T, M	Four embodied environments	1,128 vision-driven tasks from high-level household reasoning to low-level navigation and manipulation, with capability subsets.
RoboFail	Verification	Simulated and physical failure traces	Failure localization, explanation, and correction from multimodal execution history (released with REFLECT).
RoCoBench	T, M; feedback-rich	Six multi-robot manipulation tasks	Task allocation, communication, coordinated waypoint generation, and collision-aware replanning (released with RoCo).
Kitchen-R	T, M	Isaac Sim digital-twin kitchen	Planner-only, policy-only, and whole-system evaluation of language-guided mobile manipulation.
RLBench	M; I substrate	CoppeliaSim	Diverse vision-guided manipulation tasks; reused to evaluate generated programs and skills.
CALVIN	M	PyBullet	Language-conditioned long-horizon manipulation, scored by how many of five chained instructions are completed.
LIBERO	M; I substrate	robosuite / MuJoCo	Knowledge transfer across the Spatial, Object, Goal, and LIBERO-100 suites.
LIBERO-PRO	M; I substrate	robosuite / MuJoCo	Perturbed objects, initial states, instructions, and environments; used to test generated programs and persistent artifacts (CaP-X, RHO, ASPIRE).
robosuite	M; I substrate	MuJoCo	Contact-rich manipulation; reused for code-as-policy generation, controller repair, and repository optimization.
BEHAVIOR-1K	T, M; I substrate	OmniGibson	Long-horizon household activities with navigation, manipulation, articulated objects, and rich object states.
M³Bench	M	Physics-based mobile manipulation	30,000 rearrangement tasks in 119 scenes for whole-body base–arm motion generation.
RoboCasa365	T, M; I substrate	RoboCasa / MuJoCo	365 household tasks in 2,500 kitchens for multi-task, compositional, and lifelong-learning evaluation.
CaP-Bench	T, M; I substrate	CaP-Gym over robot simulators	Coding-agent performance while varying API abstraction, perceptual grounding, multi-turn interaction, visual feedback, and skill synthesis.
RAI Bench	T, M; I substrate	ROS 2 / O3DE and tool calls	Manipulation, tool-calling, and VLM benchmarks of the RAI framework; RHO optimizes a deployed agent harness on it.
ALRM Benchmark	T, M	Simulation; 54 manipulation tasks	Multistep reasoning and linguistic variation under code-as-policy and tool-as-policy execution.

The survey proposes that self-improvement benchmarks should report performance as a function of the improvement budget (physical episodes, robot-hours, human interventions, wall-clock time, or simulation interactions). They should keep development, validation, and test tasks apart, and report transfer to held-out tasks, regression on earlier capabilities, and which improvement decisions the FM made.

Open challenges

• Bridging semantic reasoning and embodiment. Performance depends on the human-designed robot interface. That interface should become adaptive, so that agents can synthesize, validate, and reuse new skills when the existing ones fall short.
• Reliable automatic verification in the physical world. A robot can improve itself only if it can tell whether a change helped. Physical critics that give calibrated outcome, progress, and diagnostic feedback remain a central bottleneck.
• Autonomous and sample-efficient physical experimentation. People still handle resets, calibration, and safety checks. Systems should also choose the most informative experiments.
• Optimizing and consolidating self-improving agents. Programs, prompts, and repositories get no gradient from physical evaluation. Better formulations should propose, compare, merge, and prune changes, and decide whether new experience stays as memory, edits a component, becomes a skill, or is distilled into a policy.
• When is more FM authority worth it? Each delegated decision widens both what the system can fix and what must be evaluated and secured. Under an equal robot and compute budget, which delegated decisions improve held-out performance enough to pay for their cost?
• Safe promotion, audit, and rollback. Improvement systems should separate experimentation from deployment, gate releases on regression and safety tests, and keep versioned artifacts and a tested rollback path.

Related surveys

★ marks the surveys that the paper compares with in detail.

Year	Paper	Venue	Scope
2026	Foundation Models in Robotics: A Comprehensive Review of Methods, Models, Datasets, Challenges and Future Research Directions ★	TMLR 2026	Organizes FMs in robotics by model type, architecture, learning paradigm, task, and domain.
2026	Self-Improvements in Modern Agentic Systems: A Survey ★	arXiv	Separates updates by a fixed external optimizer from updates an agent makes to itself; robotics is one of six domains.
2026	When Multi-Robot Systems Meet Agentic AI: Towards Embodied Collective Intelligence	arXiv	Frames collective robot intelligence around shared world, task, and skill memories.
2025	Large Model Empowered Embodied AI: A Survey on Decision-Making and Embodied Learning ★	arXiv	Covers hierarchical decision making with feedback, end-to-end VLA models, embodied learning, and world models.
2025	Towards Embodied Agentic AI: Review and Classification of LLM- and VLM-Driven Systems for Manufacturing ★	arXiv	Classifies LLM/VLM-driven robots by integration and agent role; the closest prior survey.
2025	Large Language Models for Multi-Robot Systems: A Survey	arXiv	Organizes LLM-enabled multi-robot work across communication, allocation, planning, control, interaction, and safety.
2024	Real-World Robot Applications of Foundation Models: A Review	Advanced Robotics 2024	Maps foundation models to perception, planning, control, interaction, and learning in deployed robot systems.
2024	Agent AI: Surveying the Horizons of Multimodal Interaction	arXiv	A broad multimodal-agent taxonomy connecting FM reasoning, tool use, embodiment, and multi-agent behavior.
2024	A Survey on Large Language Model Based Autonomous Agents	Frontiers of Computer Science 2024	FM-centered decision loops of software agents: profile, memory, planning, and action modules.
2024	Cognitive Architectures for Language Agents	TMLR 2024	CoALA: a framework for language agents with modular memory, structured actions, and a decision procedure.

Survey data

The search records, screening and coding protocols, decisions, and per-system labels will be released in this repository once the paper is published.

Citation

If you find this survey or list useful, please cite:

@misc{wang2026agenticrobotics,
  title  = {A Survey of Agentic Robotics: Toward Continual Self-Improvement},
  author = {Wang, Jiaming and Jiang, Yuhua and Shen, Zhengcheng and Soh, Harold},
  year   = {2026},
  url    = {https://www.alphaxiv.org/abs/2610.rsi-agentic-robotics}
}

Contributing

Contributions are welcome. See CONTRIBUTING.md for the scope, the coding rules, and the row format. The "Analyzed in the survey" tables follow the paper; please open an issue if you find an error in them.

License

This repository is available under the MIT License.