Physical RSI

Physical Recursive Self-Improvement

From adaptation to inherited improvement.

A paper-aligned, evidence-auditable map of recursive improvement in embodied systems.

Latest papers | Boundary | L0-L5 | Research map | Paper catalog

Life is not repetition, but creation. Duration carries the whole of reality: a boundless river flowing toward a horizon with no predetermined end.
_{Darwin's first evolutionary tree (1837), beside Haeckel's Pedigree of Man (1874). Both works are public domain.}

🔥 Latest Papers _top

First posted	Paper	Focus
2026-10-13	EmbodiedEvo	Reusable Skills and Lessons from robot execution
2026-10-09	RSI-Anything	World-model-guided improvement across computational and scientific tasks
2026-10-08	Embodied Turing Machines	Stateful code and shared control libraries
2026-10-07	EmbodiedRSI	Hypothesis-guided exploration and code–skill evolution
2026-10-07	Video Prediction Policy 2	Video prediction and world-action modeling
2026-10-06	PhysEvo	Execution tools, skills, and diagnostic-tool revision
2026-09-29	RoboHarn-Evo	Hierarchical physical knowledge and continued interaction

From Evolution to Physical RSI _top

Natural evolution did not become powerful because organisms changed once. Variation mattered because selection acted on it, and inheritance carried the consequences forward. Across generations, the environment became part of the history of what life could become.

Physical RSI asks whether an embodied artificial system can form an engineered counterpart to this chain. Physical interaction supplies experience, verification determines which changes are admitted, and inheritance returns accepted changes to the next improvement cycle.

Can the world change not only what a machine does, but how it produces its next improvement?

The analogy deliberately stops there. Natural selection does not owe us a protected objective, an audit trail, or a matched resource budget. An engineered recursive system does. A robot may collect more data, recover from a failed grasp, or fine-tune a policy after deployment. Retained capability gains may arise from updates to its execution system. Physical RSI additionally requires inherited revisions of the improvement procedure and independently verified gains in subsequent improvement across tasks. Grounding may come from real interaction or physics simulation; report which source drives each inherited change.

This repository is for Physical Recursive Self-Improvement research, where we study how physical interaction can drive retained changes in embodied systems and the processes that improve them. It follows the evidence boundaries of the accompanying paper and serves as a reading map, not a leaderboard: labels such as "self-improving," "autonomous," or "lifelong" are not treated as evidence on their own.

This repository is actively maintained. If an author would prefer not to have a paper listed here, please contact shangding.gu@berkeley.edu. Constructive comments, corrections, and suggestions are welcome.

You are also welcome to update the collection. To suggest a missing paper or correct an existing entry, you can:

fork this repository and open a pull request;
report an issue;
or email shangding.gu@berkeley.edu.

Content

Explore the framework, paper collection, and open questions:

🆕 Latest Papers
From Evolution to Physical RSI
The Boundary
Capability Map
Harness Adaptation, Training, and Verification
Paper Catalog
Open Questions
Citation
License

The Boundary _top

Physical RSI distinguishes the execution system M = (Θ, H)—models and their harness—from its improvement procedure Φ = (Adapt, Train, Ver). Improving task performance is not the same as improving the procedure that produces future updates.

_{Task execution, the improvement procedure, safety and control, and the L0–L5 framework.}

Capability Map _top

The framework separates what changes from where the retained gain transfers:

L0: no autonomous retained-improvement loop.
L1 / L3: inherited execution-system updates improve the same task / transfer across tasks.
L2 / L4: inherited improvement-procedure updates support same-task / cross-task improvement.
L5: successive procedure revisions produce better subsequent improvements under matched budgets and protected cross-task evaluation—not merely a changed procedure.

_{Update targets and task identity: a new scene alone does not establish a new task.}

Harness Adaptation, Training, and Verification _top

Harness adaptation: use execution feedback and improvement history to update memory, skills, tools, and workflows, retaining adopted changes for later execution.
Model training: collect and select experience, evaluate outcomes and assign credit, update model parameters, then verify and consolidate accepted changes.
Safety and control: constrain trials and reassess safeguards as the system changes.
Verification and evaluation: decide which changes may be retained, then independently measure their later benefit.
Benchmarks: assess task performance, robustness, feedback reliability, and development processes.

Paper Catalog _top

Some works span multiple functions and topics; each is listed once under the category that best reflects its primary contribution.

Browse by topic

Foundations and Surveys
- Self-Modifying Systems and Recursive Improvement
- RSI and AutoML Foundations and Surveys
Agent Harnesses and Execution
Model Training and Adaptation
Safety and Assurance
Verification and Evaluation Methods
Benchmarks, Datasets, and Platforms

Foundations and Surveys

Self-Modifying Systems and Recursive Improvement

RSI-Anything: Unlocking Recursive Self-Improvement with World Models by Taoyong Cui et al. 2026. Predictive selection, retained policy/workflow updates, and conditional world-model adaptation in computational and scientific tasks.
MetaRSI / RSI2: A Meta-Recursive Self-Improving System for Recursive Self-Improving Systems Themselves by Zihan Tan et al. 2026. Data, harness, and model operators with an editable scheduler; evaluated on digital coding and science/mathematics tasks.
Hyperagents by Jenny Zhang et al. 2026.
SelfSearch: Reward-Free Search for Self-Improving Agents by Jungwoo Yang et al. 2026.
Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents by Jenny Zhang et al. 2026.
Godel Agent: A Self-Referential Agent Framework for Recursively Self-Improvement by Xunjian Yin et al. 2025.
Self-adapting language models by Adam Zweiger et al. 2025.
Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation by Eric Zelikman et al. 2024.
Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements by Jürgen Schmidhuber. 2006.

RSI and AutoML Foundations and Surveys

Recursive Self-Improvement in AI: A Survey by Hao Wu et al. 2026.
A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence by Huan-ang Gao et al. 2026.
AI4AI survey: From long-horizon agents to recursive self-improvement—definitions, reliable horizons, and open problems by Kai Wu et al. 2026.
Diving into reliable self-evolving agents: A survey by Kaiqi Wang et al. 2026.
From AGI to ASI by Tim Genewein et al. 2026.
Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement by Hongyao Tang et al. 2026. Bibliography record; no external link supplied in v34.
Recursive self-improvement in AI: From bounded self-refinement to autonomous research loops by Mingguang Chen et al. 2026.
Self-Evolving AI for Humanoids: Mechanisms, Safety, and Evaluation of Post-Deployment Self-Improvement by Loc X. Nguyen et al. 2026.
Self-improvements in modern agentic systems: A survey by Zhe Ren et al. 2026.
Self-improving agents in the era of experience: A survey of self- to meta-evolution by Che Jiang et al. 2026.
The last AI built by humans: Toward genuine recursive self-improvement by Yi Duan et al. 2026.
The path to recursive self-improving agents: Foundation, framework, and future directions by Shuaiqi Liu et al. 2026.
A comprehensive survey of self-evolving AI agents: A new paradigm bridging foundation models and lifelong agentic systems by Jinyuan Fang et al. 2025.
AutoML: A Survey of the State-of-the-Art by Xin He et al. 2021.
Automated Machine Learning: Methods, Systems, Challenges by Frank Hutter et al. 2019.
Towards AI That Improves Itself: A Survey of Recursive Self-Improvement by Hanjing Li et al. n.d.

Agent Harnesses and Execution

Memory, skills, tools, and workflows used during task execution, together with methods for updating these resources.

Memory and Knowledge Management

Includes persistent execution guidance and tool documentation, as well as explicit memory systems.

RoboHarn-Evo: Evolving Hierarchical Physical Knowledge for Self-Improving Robotic Manipulation by Shifeng Bao et al. 2026. Task- and action-level knowledge is revised from interaction and reused on held-out configurations.
Beyond Static Toolsets: Self-Evolving LLM Tool Agents via Continual Documentation Adaptation by Bin Wu et al. 2026.
HomeBody: A humanoid that explores, remembers, and acts on its own by Gio Huh et al. 2026.
Know Your Body: A Harness for Direct and Self-Improving Robot Control with VLMs by Zeyu Lou et al. 2026.
PhysMem: Scaling Test-Time Memory for Embodied Physical Reasoning by Haoyang Li et al. 2026.
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution by Liyan Tang et al. 2026.
Learning to Continually Learn via Meta-learning Agentic Memory Designs by Yiming Xiong et al. 2026.
MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents by Haozhen Zhang et al. 2026.
DynaMem: Online Dynamic Spatio-Semantic Memory for Open World Mobile Manipulation by Peiqi Liu et al. 2025.
MemEvolve: Meta-Evolution of Agent Memory Systems by Guibin Zhang et al. 2025.
AutoGuide: Automated Generation and Selection of Context-Aware Guidelines for Large Language Model Agents by Yao Fu et al. 2024.
AutoManual: Constructing Instruction Manuals by LLM Agents via Interactive Environmental Learning by Minghao Chen et al. 2024.
Distilling and Retrieving Generalizable Knowledge for Robot Manipulation via Language Corrections by Lihan Zha et al. 2024.
ExpeL: LLM Agents Are Experiential Learners by Andrew Zhao et al. 2024.
Reflexion: Language Agents with Verbal Reinforcement Learning by Noah Shinn et al. 2023.

Skill Libraries and Tool Construction

EmbodiedEvo: Self-Evolving Coding Agent For Physical Intelligence by Wenyi Li et al. 2026. Freezes Playground-acquired code Skills and diagnostic Lessons for fresh agents in unseen RoboDojo-Sim scenes.

Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement by Kairui Hu et al. 2026. Method: Code-Only-as-Policy (COAP); coding agents improve a shared library offline, with fixed code controlling runtime execution.
Act-Observe-Rewrite: Multimodal Coding Agents as In-Context Policy Learners for Robot Manipulation by Vaishak Kumar. 2026.
ASPIRE: Agentic /Skills Discovery for Robotics by Runyu Lu et al. 2026.
Dynamo: Dynamic Skill-Tool Evolution for Vision-Language Agents by Yutao Sun et al. 2026.
EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents by Ruofei Ju et al. 2026.
EvoSkill: Automated Skill Discovery for Multi-Agent Systems by Salaheddin Alzubi et al. 2026.
Playful Agentic Robot Learning by Junyi Zhang et al. 2026.
RACaP: Agentic Reasoning, Acting, and Coding as Policies for Evolvable Robot Learning by Zexi Li et al. 2026.
Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents by Yen-Jen Wang et al. 2026.
SkillAdam: Stable and Efficient Skill Evolution for Agents by Gaoyuan Li et al. 2026.
Meta Context Engineering via Agentic Skill Evolution by Haoran Ye et al. 2026.
SkillOpt: Executive Strategy for Self-Evolving Agent Skills by Yifan Yang et al. 2026.
SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills by Boyuan Zheng et al. 2025.
ViReSkill: Vision-Grounded Replanning with Skill Memory for LLM-Based Planning in Lifelong Robot Learning by Tomoyuki Kagaya et al. 2025.
CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets by Lifan Yuan et al. 2024.
Lifelong Robot Library Learning: Bootstrapping Composable and Generalizable Skills for Embodied Control with Language Models by Georgios Tziafas and Hamidreza Kasaei. 2024.
ReGAL: Refactoring Programs to Discover Generalizable Abstractions by Elias Stengel-Eskin et al. 2024.
Large Language Models as Tool Makers by Tianle Cai et al. 2024.
CREATOR: Tool Creation for Disentangling Abstract and Concrete Reasoning of Large Language Models by Cheng Qian et al. 2023.
SkiROS2: A skill-based Robot Control Platform for ROS by Matthias Mayr et al. 2023.
Code as Policies: Language Model Programs for Embodied Control by Jacky Liang et al. 2023.

Planning and Workflows

EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution by Python Song et al. 2026. Hypothesis graphs, value-of-information experiment selection, code–skill co-evolution, and reward-grounded memory management.

PhysEvo: Astra Can Act, Let It by Wenqing Tian et al. 2026. Retained revisions to execution tools, skills, and meta-agent diagnostic utilities around a frozen model.
Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence by Hongcheng Gao et al. 2026. Method: Physical Coding / HexaAnything; separate harness, tool, and initial trace-to-model improvements from the broader co-evolution roadmap.
AdaHVLA: Adaptive Harnesses for Long-Horizon Vision-Language-Action Execution by Junyi Tang et al. 2026. Retains coordination-code revisions and their supporting evidence; changing the execution harness does not alone establish a better improvement procedure.
Show-Harness: Just a VLM Agent Can Play Robots by Yanzhe Chen et al. 2026. Semantic robot-control interfaces and GUI-based demonstration collection.
Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence by Xin Ding et al. 2026. Runtime critics, recovery skills, and validation-gated harness updates around a frozen policy.
RoboHarness: Memory-Driven Orchestration of Heterogeneous Robot Policies for Long-Horizon Planning by Jinbang Huang et al. 2026. Capability-aware policy routing, Memory Bridge handoffs, and execution-log-driven refinement.
Don’t Drop the BATON: Long-Horizon Robot Manipulation via Agentic Subtask Exploration and Transition-aware Memory by Bingxin Xu et al. 2026.
Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents by Yixian Zhang et al. 2026.
HarnessPAI: An Evolving Harness for Physical AI by Xin Wang et al. 2026.
Kintsugi: Learning Policies by Repairing Executable Knowledge Bases by Teng Cao et al. 2026.
Living-Harness Is an Interactive-Agent Evolver by Yuetian Du et al. 2026.
RHO: Your Coding Agent is Secretly a Roboticist by Karim Elmaaroufi et al. 2026.
RoboFoundry: System-as-Policy Evolution for Self-Learning Embodied Agents by Jingsong Liang et al. 2026.
Self-Harness: Harnesses That Improve Themselves by Hangfan Zhang et al. 2026.
Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses by Tailin Zhou. 2026.
VERSE: Verified Self-Evolving Optimizer for Agent Harnesses by Zekai Wang et al. 2026.
RegenHarness: A Robot Agent Harness with Evidence-Gated Recursive Self-Improvement by Kailin Wang et al. 2026.
AutoRT: Embodied Foundation Models for Large Scale Orchestration of Robotic Agents by Michael Ahn et al. 2024.
Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution by Chrisantha Fernando et al. 2024.
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances by Michael Ahn et al. 2022.
Behavior Trees in Robotics and AI: An Introduction by Michele Colledanchise and Petter Ögren. 2018.

Failure Diagnosis and Execution Recovery

From Failed Trajectories to Reliable LLM Agents: Diagnosing and Repairing Harness Flaws by Mengzhuo Chen et al. 2026.
DoReMi: Grounding Language Model by Detecting and Recovering from Plan-Execution Misalignment by Yanjiang Guo et al. 2024.

Model Training and Adaptation

Data preparation, learning methods, and learned models for robot improvement, including human-assisted learning, autonomous practice, and continual transfer.

Self-Supervised and Representation Learning

Video Prediction Policy 2: Predict Better, Act Better by Yanjiang Guo et al. 2026. Method: VPP2; manipulation-video pretraining, visual-planner distillation, and world-action modeling.
RoboTTT: Context Scaling for Robot Policies by Yunfan Jiang et al. 2026.
Self-Improving Loops for Visual Robotic Planning by Calvin Luo et al. 2025. Method: SILVR; inspected version: v3. Successful self-produced trajectories fine-tune a video planner; the improvement recipe is not thereby shown to evolve.
Latent action pretraining from videos by Seonghyeon Ye et al. 2025.
Learning interactive real-world simulators by Sherry Yang et al. 2024.
LIV: Language-Image Representations and Rewards for Robotic Control by Yecheng Jason Ma et al. 2023.
Structured world models from human videos by Russell Mendonca et al. 2023.
Real-world robot learning with masked visual pre-training by Ilija Radosavovic et al. 2022.

Continual Adaptation and Knowledge Transfer

Can vision-language-action models learn from real-world data continually without forgetting? by Jiarun Zhu et al. 2026.
EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation by Haoran Lang et al. 2026.
F4R: Failure-Driven Recognition, Reconstruction, Refinement, and Redeployment for Continual Robot Self-Improvement by Zhuoyuan Yu et al. 2026.
Pretrained Vision-Language-Action Models Are Surprisingly Resistant to Forgetting in Continual Learning by Huihan Liu et al. 2026.
Robust Finetuning of Vision-Language-Action Robot Policies via Parameter Merging by Yajat Yadav et al. 2026.
Towards human-like physical intelligence: Lifelong vision-language-action learning for robotic manipulation by Yao He et al. 2026.
EXTRACT: Efficient Policy Learning by Extracting Transferable Robot Skills from Offline Data by Jesse Zhang et al. 2025.
Knowledge diversion for efficient morphology control and policy transfer by Fu Feng et al. 2025.
SRSA: Skill Retrieval and Adaptation for Robotic Assembly Tasks by Yijie Guo et al. 2025.
LOTUS: Continual Imitation Learning for Robot Manipulation through Unsupervised Skill Discovery by Weikang Wan et al. 2024.
PRISE: LLM-Style Sequence Compression for Learning Temporal Action Abstractions in Control by Ruijie Zheng et al. 2024.

Training Procedures and Algorithm Optimization

Differentiable evolutionary reinforcement learning by Sitao Cheng et al. 2026. Method: DERL.
ENPIRE: Agentic Robot Policy Self-Improvement in the Real World by Wenli Xiao et al. 2026.
EvoTrainer: Co-Evolving LLM Policies and Training Harnesses for Autonomous Agentic Reinforcement Learning by Guhong Chen et al. 2026.
HARBOR: A Harness Framework for Agentic Robot Reinforcement Learning by Zechu Li et al. 2026.
AlphaEvolve: A coding agent for scientific and algorithmic discovery by Alexander Novikov et al. 2025.
Discovering State-of-the-Art Reinforcement Learning Algorithms by Junhyuk Oh et al. 2025. Method: DiscoRL.
Eureka: Human-Level Reward Design via Coding Large Language Models by Yecheng Jason Ma et al. 2024.
Automated Self-Supervised Learning for Graphs by Wei Jin et al. 2022. Method: AutoSSL.
Auto-sklearn 2.0: Hands-free AutoML via Meta-learning by Matthias Feurer et al. 2022. Bibliography record; no external link supplied in v34.

Robot Learning Surveys and Foundations

Training Methods and Generation Technologies for Embodied Intelligent Robot Manipulation Skill Models: A Systematic Review by Lianpeng Li et al. 2026.
World model for robot learning: A comprehensive survey by Bohan Hou et al. 2026.
Vision-Language-Action Models for Robotics: A Review towards Real-World Applications by Kento Kawaharazuka et al. 2025.
A comprehensive survey of continual learning: Theory, method and application by Liyuan Wang et al. 2024.
Continual Learning for Robotics: Definition, Framework, Learning Strategies, Opportunities and Challenges by Timothée Lesort et al. 2020.
The ingredients of real-world robotic reinforcement learning by Henry Zhu et al. 2020.
Challenges of real-world reinforcement learning by Gabriel Dulac-Arnold et al. 2019.
Continual Lifelong Learning with Neural Networks: A Review by German I. Parisi et al. 2019.

Safety and Assurance

Safety during learning, acceptance, and deployment, including runtime safeguards, oversight, and integrity of the improvement process.

Safe Exploration and Constrained Learning

PACT: Self-Evolving Physical Safety Alignment for Diffusion Policies in Embodied Manipulation by Lingxuan Wu et al. 2026. Constraint-gradient post-training of diffusion policies with a progressively tightened safety curriculum.
CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions by Lizhi Yang et al. 2026.
Safe Continual Reinforcement Learning in Non-stationary Environments by Austin Coursey et al. 2026.
ROSCOM: Robust Safe Reinforcement Learning on Stochastic Constraint Manifolds by Shangding Gu et al. 2025.
Safe and Balanced: A Framework for Constrained Multi-Objective Reinforcement Learning by Shangding Gu et al. 2025.
Safe Continual Domain Adaptation after Sim2Real Transfer of Reinforcement Learning Policies in Robotics by Josip Josifovski et al. 2025.
SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning by Borong Zhang et al. 2025.
Balance Reward and Safety Optimization for Safe Reinforcement Learning: A Perspective of Gradient Manipulation by Shangding Gu et al. 2024.
Safe Multi-Agent Reinforcement Learning for Multi-Robot Control by Shangding Gu et al. 2023. Benchmarks: Safe MAMuJoCo, Safe MARobosuite, Safe MAIG.
COptiDICE: Offline Constrained Reinforcement Learning via Stationary Distribution Correction Estimation by Jongmin Lee et al. 2022.
End-to-End Safe Reinforcement Learning through Barrier Functions for Safety-Critical Continuous Control Tasks by Richard Cheng et al. 2019.
Safe Exploration in Continuous Action Spaces by Gal Dalal et al. 2018.
Constrained Policy Optimization by Joshua Achiam et al. 2017.
Safe Modelbased Reinforcement Learning with Stability Guarantees by Felix Berkenkamp et al. 2017.
Safe Controller Optimization for Quadrotors with Gaussian Processes by Felix Berkenkamp et al. 2016.
Safe Exploration in Finite Markov Decision Processes with Gaussian Processes by Matteo Turchetta et al. 2016.
Risk-Sensitive and Robust Decision-Making: a CVaR Optimization Approach by Yinlam Chow et al. 2015.

Safe Transfer and Update Acceptance

Robust Transfer of Safety-Constrained Reinforcement Learning Agents by Markel Zubia et al. 2025.
Sim-to-Lab-to-Real: Safe Reinforcement Learning with Shielding and Generalization Guarantees by Kai-Chieh Hsu et al. 2023.
Towards Safe Policy Improvement for Non-Stationary MDPs by Yash Chandak et al. 2020.
Preventing Undesirable Behavior of Intelligent Machines by Philip S. Thomas et al. 2019.
Safe Policy Improvement with Baseline Bootstrapping by Romain Laroche et al. 2019.
Verifiably Safe Off-Model Reinforcement Learning by Nathan Fulton and André Platzer. 2019.
High Confidence Policy Improvement by Philip S. Thomas et al. 2015.

Runtime Control and Failure Monitoring

Failure Identification in Imitation Learning Via Statistical and Semantic Filtering by Quentin Rolland et al. 2026.
Code-as-Monitor: Constraint-Aware Visual Programming for Reactive and Proactive Robotic Failure Detection by Enshen Zhou et al. 2025.
Failure Prediction at Runtime for Generative Robot Policies by Ralf Römer et al. 2025.
Realizable Continuous-Space Shields for Safe Reinforcement Learning by Kyungmin Kim et al. 2025.
SAFE: Multitask Failure Detection for Vision-Language-Action Models by Qiao Gu et al. 2025.
Unpacking Failure Modes of Generative Policies: Runtime Monitoring of Consistency and Progress by Christopher Agia et al. 2025.
Evaluating Uncertainty-Based Failure Detection for Closed-Loop LLM Planners by Zhi Zheng et al. 2024.
Failure Prediction with Statistical Guarantees for Vision-Based Robot Control by Alec Farid et al. 2022.
A Predictive Safety Filter for Learning-Based Control of Constrained Nonlinear Dynamical Systems by Kim Peter Wabersich and Melanie N. Zeilinger. 2021.
A Predictive Safety Filter for Learning-Based Racing Control by Ben Tearle et al. 2021.
Measurement-Robust Control Barrier Functions: Certainty in Safety with Uncertainty in State by Ryan K. Cosner et al. 2021.
Recovery RL: Safe Reinforcement Learning With Learned Recovery Zones by Brijen Thananjeyan et al. 2021.
Neural Simplex Architecture by Dung T. Phan et al. 2020.
A General Safety Framework for Learning-Based Control in Uncertain Robotic Systems by Jaime F. Fisac et al. 2019.
SOTER: A Runtime Assurance Framework for Programming Safe Robotics Systems by Ankush Desai et al. 2019.
Leave no trace: Learning to reset for safe and autonomous reinforcement learning by Benjamin Eysenbach et al. 2018.
Safe reinforcement learning via shielding by Mohammed Alshiekh et al. 2018.
Control Barrier Function Based Quadratic Programs for Safety Critical Systems by Aaron D. Ames et al. 2017.
Safety Barrier Certificates for Collisions-Free Multirobot Systems by Li Wang et al. 2017.

Oversight and Integrity Protection

Reward Hacking Challenges Oversight of Autonomous Research Agents by Yue Huang et al. 2026.
Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection by Kai Greshake et al. 2023.
Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners by Allen Z. Ren et al. 2023.
The Off-Switch Game by Dylan Hadfield-Menell et al. 2017.
Safely Interruptible Agents by Laurent Orseau and Stuart Armstrong. 2016.

Safety Surveys and Risk Taxonomies

Evolutionary safety of recursive self-improving AI: Taxonomy, risk discovery, and evaluation by Chang Gong et al. 2026.
A Review of Safe Reinforcement Learning: Methods, Theories, and Applications by Shangding Gu et al. 2024.

Verification and Evaluation Methods

Candidate verification, verifier reliability, and independent evaluation of improvement outcomes.

Verification of Candidate Updates

Visual Verification Enables Inference-Time Steering and Autonomous Policy Improvement by Mingtong Zhang and Dhruv Shah. 2026.

Verifier Reliability and Meta-Verification

What Stops Recursive Self-Improvement in Robotics? Lessons from 123 Rounds of Agentic Skill Discovery by Jiaming Wang. 2026.
No Free Checker: A Survey of Verifiers for Robot Policies by Yang Wan et al. 2026.

Evaluation of Improvement Outcomes

A survey on evaluation of embodied AI by Liyu Hou et al. 2026.
Beyond Binary Success: Sample-Efficient and Statistically Rigorous Robot Policy Comparison by David Snyder et al. 2026.
Is Your Imitation Learning Policy Better than Mine? Policy Comparison with Near-Optimal Stopping by David Snyder et al. 2025.
Doubly Robust Off-policy Value Evaluation for Reinforcement Learning by Nan Jiang and Lihong Li. 2016.
Generalization in Adaptive Data Analysis and Holdout Reuse by Cynthia Dwork et al. 2015.

Benchmarks, Datasets, and Platforms

Datasets, task suites, evaluation resources, and benchmarks for embodied systems and research agents.

Shared Robot Datasets

DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset by Alexander Khazatsky et al. 2024.
Open X-Embodiment: Robotic Learning Datasets and RT-X Models by Open X-Embodiment Collaboration. 2024.

Task Learning, Performance, and Generalization

RoboQuest: Generalist Physical Agents that Search, Inspect and Test by Liu Renhang et al. 2026. Tests goal-directed search, inspection, and interactive testing within an episode; cross-episode inheritance requires an additional protocol.
EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks by Lizhou Liang et al. 2026. Tests memory-conditioned action and experience generalization; it does not alone show improvement of the memory-construction procedure.
RoboCasa365: A Large-Scale Simulation Framework for Training and Benchmarking Generalist Robots by Soroush Nasiriany et al. 2026.
EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents by Rui Yang et al. 2025.
ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI by Stone Tao et al. 2025.
Towards Generalizable Vision-Language Robotic Manipulation: A Benchmark and LLM-guided 3D Policy by Ricardo Garcia et al. 2025. Benchmark: GemBench.
VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks by Shiduo Zhang et al. 2025.
RoboTwin: Dual-Arm Robot Benchmark with Generative Digital Twins by Yao Mu et al. 2025.
BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation by Chengshu Li et al. 2024.
LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning by Bo Liu et al. 2023.
VIMA: Robot Manipulation with Multimodal Prompts by Yunfan Jiang et al. 2023. Benchmark: VIMA-Bench.
CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks by Oier Mees et al. 2022.
What Matters in Learning from Offline Human Demonstrations for Robot Manipulation by Ajay Mandlekar et al. 2021. Resource: robomimic.
Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning by Tianhe Yu et al. 2020.
robosuite: A Modular Simulation Framework and Benchmark for Robot Learning by Yuke Zhu et al. 2020.
RLBench: The Robot Learning Benchmark & Learning Environment by Stephen James et al. 2019.

Robustness and Physical-World Evaluation

EmbodiedGovBench: A Benchmark for Governance, Recovery, and Upgrade Safety in Embodied Agent Systems by Xue Qin et al. 2026. Governance measurement framework with prototype instantiations; released artifacts and empirical coverage require separate checks.
RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation by Tianxing Chen et al. 2026.
VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models by Borong Zhang et al. 2026.
RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies by Tianxing Chen et al. 2026.
LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models by Senyu Fei et al. 2025.
LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization by X. Zhou et al. 2025.
REALM: A Real-to-Sim Validated Benchmark for Generalization in Robotic Manipulation by Martin Sedlacek et al. 2025.
AutoEval: Autonomous Evaluation of Generalist Robot Manipulation Policies in the Real World by Zhiyuan Zhou et al. 2025.
RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies by Pranav Atreya et al. 2025.
AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions by Zonghao Ying et al. 2025. Simulated safety assessment across perception, planning, and execution; distinct from the SAFE failure detector in Runtime Control and Failure Monitoring.
Evaluating Real-World Robot Manipulation Policies in Simulation by Xuanlin Li et al. 2024. Benchmark: SIMPLER.
THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulation by Wilbert Pumacay et al. 2024.

Feedback and Evaluation Reliability

FailBench: How Reliable Are VLMs at Judging Robot Task Success? by Zaruhi Navasardyan et al. 2026.
Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons by Anthony Liang et al. 2026. Benchmark: RBM-EVAL.
RoboReward: General-Purpose Vision-Language Reward Models for Robotics by Tony Lee et al. 2026. Benchmark: RoboRewardBench.
Scaling Robot Policy Evaluation via Discrete Diffusion World Model by Yaxuan Li et al. 2026. Resource: dWorldEval.
SC3-Eval: Evaluating Robot Foundation Models via Self-Consistent Video Generation by Wei-Cheng Tseng et al. 2026.
WorldEval: World Model as Real-World Robot Policies Evaluator by Yaxuan Li et al. 2025.

Robot Engineering and Development

Embodied-BenchForge: A Closed-Loop Agentic Workflow for Embodied Benchmark Construction by Baoyang Jiang et al. 2026. Constructs and repairs benchmark artifacts with dependency tracking; development-time verification is distinct from protected final evaluation.
CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation by Letian Fu et al. 2026. Benchmark: CaP-Bench (CaP-X).
From Digital to Physical: Digital Agents as Autonomous Coaches for Physical Intelligence by Zixing Lei et al. 2026. Benchmark: RoboCoach-Bench.
RLE-Bench: A Qualifying Exam for Coding Agents as Robot Learning Engineers by Haitong Ma et al. 2026.

RSI related Benchmarks

AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement by Yizhe Chi et al. 2026. Benchmark: AI4AI-Bench.
HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness? by Yuhao Wu et al. 2026.
OpenRSI Index by OpenRSI Foundation. 2026.
PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents by Shuhan Xue et al. 2026.
RSI-Exam: Benchmarking Recursive Self-Improvement through Executable Research by RSI-Exam Team. 2026.
RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement by Fanqing Meng et al. 2026.
RSIGym: A Flexible Environment for Recursive Self-Improvement by Fanqing Meng et al. 2026.
MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks by Zexue He et al. 2026.
Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions by Yuanzhe Hu et al. 2026. Benchmark: MemoryAgentBench.
MLGym: A New Framework and Benchmark for Advancing AI Research Agents by Deepak Nathani et al. 2025. Benchmark: MLGym-Bench.
MLRC-Bench: Can Language Agents Solve Machine Learning Research Challenges? by Yunxiang Zhang et al. 2025.
RE-Bench: Evaluating Frontier AI R&D Capabilities of Language Model Agents against Human Experts by Hjalmar Wijk et al. 2025.
LifelongAgentBench: Evaluating LLM Agents as Lifelong Learners by Junhao Zheng et al. 2025.
MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation by Qian Huang et al. 2024.
MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering by Jun Shern Chan et al. 2024.

Open Questions _top

Section 7 of the v35 manuscript identifies six open problems, listed in manuscript order:

Learning from uncertain physical experience: derive reliable learning signals from partial, noisy, and stochastic feedback without reinforcing errors through inherited training-procedure changes.
Causal attribution across harness components: distinguish failures in memory, skills, and workflows from their interactions, and isolate the effects of concurrent edits under limited physical trial budgets.
Non-stationarity in the improvement loop: track changing component dependencies and physical conditions, and reassess past evidence and improvement strategies without excessive revalidation costs.
Reward hacking and evaluation leakage: independently check physical outcomes, control evaluation feedback, and use fresh or statistically corrected final tests to distinguish real gains from exploitation of imperfect judgments.
Safety across recursive updates: selectively revalidate assumptions and safeguards as procedures evolve, preserve protected requirements and effective interruption, and account for risk from rejected trials, resets, and recovery.
Benchmarking verified recursive gain: track autonomous procedure revisions, inheritance, and reuse, and compare successive procedures from the same starting system and experience under matched tasks and resource budgets.

These are gaps scoped to the reviewed manuscript, not claims that no related work exists elsewhere.

Citation

If you find the repository useful, please cite the study

@misc{zhou2026physicalrsi,
  title         = {{Who Improves the Robot? Toward Physical Recursive Self-Improvement}},
  author        = {Hai Zhou and Fan Yang and Zeyi Chen and Lejun Zhang
                   and Chenyang Fu and Jianping Li and Jiayi Zhu
                   and Xianyang Liu and Yingxuan Yang and Lili Meng
                   and Lu Wang and Xinyang Han and Jiankai Sun
                   and Weirui Ye and Guanchu Wang and Wanhua Li
                   and Ming Zhou and Boyi Li and Ming Jin and Kun Shao
                   and Muhao Chen and Guang Chen and Lei Ma
                   and Joey Tianyi Zhou and Costas Spanos and Xia Hu
                   and Mac Schwager and Weinan Zhang and Qian Feng
                   and Shangding Gu},
  year          = {2026},
  eprint        = {2610.18842},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2610.18842}
}

License _top

This repository is released under the MIT License. Linked papers and third-party materials remain subject to their respective licenses.

Name		Name	Last commit message	Last commit date
docs		docs
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md

Latest commit

History

Folders and files

Repository files navigation

Physical RSI

🔥 Latest Papers top

From Evolution to Physical RSI top

The Boundary top

Capability Map top

Harness Adaptation, Training, and Verification top

Paper Catalog top