RoboDoctor: Training-Free Recursive Self-Improvement for Robot Policies

Heyang Yu; Haomin Li; Sihang Li; Jixin Yan; Shiyun Peng; Wanqi Zhong; Tianyu Xin; Yuan Yao

cs.ROOct 26, 2026

RoboDoctor: Training-Free Recursive Self-Improvement for Robot Policies

Authors: Heyang Yu, Haomin Li, Sihang Li, Jixin Yan, Shiyun Peng, Wanqi Zhong, Tianyu Xin, Yuan Yao,

Organizations: Tsinghua University · New York University · Northwestern University

Abstract

Generalist robot policies and language-model agents provide complementary capabilities, but completing a task also requires adapting their use to the consequences of physical interaction. We study execution as a locus of adaptation: feedback guides the current action and supplies evidence for revising subsequent decisions under fixed model weights. We introduce RoboDoctor, a training-free framework that connects diagnosis, grounded action, and persistent experience across three revision targets. The harness allocates control between policy and agent; intervention determines when to assist while preserving useful policy behavior; and operation refines physical methods through observed effects. Linked experience records retain the reason for a decision, the attempted method, its outcome, and conditions for reuse. On a curated archive of 54 RoboDojo-Sim task entries, RoboDoctor achieves a five-dimension mean score of 57.10, with a native success rate of 49.44%. Recorded interactions illustrate strategy changes, selective correction, and grasp revision. These findings characterize execution-time adaptation in the evaluated system; controlled trials remain necessary to isolate the contribution of cross-episode experience.

Figures & tables

Figure 1: RoboDoctor: training-free adaptation across three levels of execution. Harness RSI allocates control between the policy and the Doctor; Intervention RSI decides when to retain or correct policy motion; Operation RSI revises physical methods and verifies their effects. External experience connects these decisions while both models remain frozen. The radar chart summarizes five capability scores for Warmup, Refinement, and the published GPT-6-Astra reference. Recorded examples illustrate alignment correction followed by pouring, and grasp revision followed by a test lift.

Figure 2: RoboDoctor architecture: experience-guided control and grounded execution. After three HYBRID Warmup episodes, Harness RSI selects HYBRID or DIRECT control and keeps this choice fixed. In HYBRID, the Doctor reviews a WAM proposal and its robot-only forward-kinematics preview, then retains up to 16 unchanged policy steps or issues a grounded correction of 1–5 physical steps. Corrections and DIRECT actions ground the current scene, execute a physical method, and verify its observed effect. The illustrated pouring episode separates a lateral alignment correction that preserves the grasp from subsequent policy-led pouring. Task experience links diagnostic cues, attempted operations, and observed outcomes to inform later decisions.

Method	Source	Score $\uparrow$	Success (%) $\uparrow$	Episodes
HexaShell-Preview	Published	53.18	46.85	—
WorldScape-R 2 SI	Published	46.58	39.35	—
EmbodiedEvo	Published	45.08	38.95	—
VPP2	Published	39.26	32.26	—
PhysicalRSI	Published	36.27	31.38	—
GPT-6-Astra	Published	28.97	22.48	—

Table 1: Overall performance on RoboDojo-Sim. Scores and native success rates follow the benchmark’s five-dimension aggregation across 54 task entries. Reference values are published leaderboard results, while RoboDoctor values summarize selected local outcomes. Evaluation protocols and interaction budgets are not matched.

Method	Gen.	Precision	Long-Hor.	Memory	Open
HexaShell-Preview	39.22	41.25	42.91	84.53	57.97
WorldScape-R 2 SI	42.95	23.51	51.48	65.78	49.16
EmbodiedEvo	35.22	25.15	33.83	79.40	51.81
VPP2	36.38	43.03	56.28	56.01	4.61
PhysicalRSI	21.30	38.05	46.37	46.74	28.88
GPT-6-Astra	33.36	12.65	21.45	43.04	34.36

Table 2: Performance across capability dimensions. Gen. denotes the equally weighted mean of the standard and randomized Generalization scores. The overall score is the equally weighted mean of the five capability scores. Reference rows report published values; the final two rows summarize selected local outcomes.

Figure 3: Aggregate performance and variation across tasks. (a) Overall scores for RoboDoctor during Warmup and Refinement and for published reference methods, using the equally weighted mean of five capability scores. (b) Changes in task-level mean score from Warmup to Refinement across all 54 entries, ordered from the largest decrease to the largest increase: 24 improve, 20 decline, and 10 remain unchanged. Annotations highlight representative gains and a regression. RoboDoctor scores summarize selected local outcomes; reference methods use different evaluation protocols, so the comparison does not isolate the effects of individual components.

Figure 4: Recorded execution changes across three RSI levels. (a) Harness RSI: an unresolved Warmup attempt is contrasted with supported tool pushing in a later DIRECT episode; the dashed divider separates episodes. (b) Intervention RSI: the Doctor corrects bottle–cup alignment while preserving the grasp, after which the WAM pours and withdraws the bottle. (c) Operation RSI: the Doctor revises a slipped grasp and verifies retention with a test lift before WAM-led placement. Blue labels indicate WAM actions or outcomes; purple labels indicate Doctor actions, outcomes, or diagnosis. Frames show selected moments within each episode.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Evidence	What it supports	What it does not establish
Policy proposal / FK preview	Intended joint targets and predicted robot poses	Actual tracking, object motion, contact, or task success
Measured end-effector pose	Achieved robot pose and target error	Stable grasp or correct placement
Fresh RGB-D interaction	Observed contact, support, transfer, or visible failure	Hidden quantitative predicates or full-task completion
Native outcome	Evaluator-defined score, validity, and success	The causal benefit of one intervention or memory entry

Appendix

Table 3: Evidence contracts for grounded action. These checks distinguish levels of support; none replaces the final native task outcome.

Task	Route	W	R	$\Delta$	SR	$n$
align_blocks	Direct	0.00	100.00	+100.00	100.00	6/6
arrange_largest_number	Hybrid	76.67	63.33	-13.33	50.00	6/6
arrange_largest_number_random	Hybrid	48.33	61.67	+13.33	50.00	6/6
build_tower	Hybrid	30.00	100.00	+70.00	100.00	3/6
classify_objects	Hybrid	100.00	80.00	-20.00	66.67	6/6
classify_objects_by_language	Direct	26.67	43.33	+16.67	16.67	6/6

Appendix

Table 4: All 54 task entries. W: Warmup score; R: Refinement score;

\Delta=\mathrm{R}-\mathrm{W}

; SR: Refinement success rate (%);

n

: valid Refinement episodes out of six nominal slots. All tasks have three Warmup outcomes. Route describes the selected Refinement records; Mixed identifies retrospective combination of both modes. Native task IDs are preserved.

Task / panel	Visible evidence	Supported interpretation
Align Blocks, Figure 4 (a)	Unaligned blocks; two-point supported pushing; aligned blocks	The later Direct episode uses a supported pushing strategy to align the blocks.
Pour Liquid into Cup, Figure 4 (b)	Lateral offset; correction with the grasp retained; pouring	Local alignment is corrected while preserving the grasp, followed by policy-driven pouring.
Build Tower, Figure 4 (c)	Slippage; end-face grasp; test lift; placement	Revised contact is checked through a lift before the block is placed.

Appendix

Table 5: Visual evidence in Figure 4 . Each example connects a revised interaction to its depicted outcome.

Robot Policy Adaptation

Robot policy adaptation changes or steers learned robot behaviors to handle new tasks, environments, hardware, and user preferences.

See all 170 papers

EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation
Sep 30, 2026Haoran Lang, Haotao Lu, Shiyu Sang +6Robot Policy Adaptation Robotic Data Collection
InstAdapt · ShanghaiTech University
Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents
Oct 1, 2026Yen-Jen Wang, Haozhe Jiang, Shuying Deng +7Robot Skill Learning Large Language Model-Based Robot Planning
UC Berkeley · Amazon FAR · MIT +1
What Does It Take For A Generalist Robot Policy To Adapt Using Minimal Data?
Aug 11, 2026Shreyas Kowshik, Sreyas Venkataraman, Leo Wang +3RL for Robotics Robot Policy Adaptation
Carnegie Mellon University

Sep 30, 2026J/K move · Enter open · S save

EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation

Haoran Lang, Haotao Lu, Shiyu Sang, Haoyang Luo+5

InstAdapt · ShanghaiTech University

Robot Policy Adaptation Robotic Data Collection Robotics Simulation Sim-to-Real Transfer Data-Efficient Robot Learning Recursive Self-Improvement

Open paper AskPDF arXiv

Adapting robot manipulation policies to new tasks and environments remains highly data-intensive, while the data needed for further improvement depends on the policy's current capabilities and failure modes. We introduce EmbodiRSI, an agentic system for recursive self-improvement (RSI) in a real-to-sim-to-real setting, where task-specific simulations are constructed from target deployment scenarios and used as low-cost environments for iterative policy improvement before transfer back to the physical world. EmbodiRSI uses policy execution feedback to guide subsequent experience acquisition and policy updates. Two complementary mechanisms close this loop: Collaborative Error Correction generates agent-assisted corrective trajectories from policy-reached states, while Adaptive Data Collection directs expert demonstration generation toward the current policy's weaknesses. The task-specific simulation serves as a reusable workspace for policy warm-up, repeatable evaluation, failure diagnosis, and targeted data generation across successive RSI rounds. Across three tabletop environments and 14 subtasks, EmbodiRSI increases scene-balanced autonomous simulation success from 50.4% to 83.5% over two RSI updates. With 400 adaptive simulated trajectories and only ten real-world refinement trajectories per subtask, EmbodiRSI achieves 83.1% scene-balanced autonomous real-world success, compared with 75.0% for adaptation using 200 real-world demonstrations per subtask. These results demonstrate that feedback-driven recursive improvement in deployment-specific simulations can enable data-efficient adaptation of embodied policies to physical environments.

Abstract

Robot Policy Adaptation

Explore similar work

EmbodiRSI: Recursive Self-Improvement for Data-Efficient Robot Adaptation

Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents

What Does It Take For A Generalist Robot Policy To Adapt Using Minimal Data?