acceptodds
Under review as a conference paper at ICLR 2027

EmbodiedEvo: Continual Robot Learning through Hypothesis-Guided Co-Evolution

Abstract

Robot foundation models provide strong visuomotor control, yet their performance can degrade when object positions or task instructions change. Further improvements often require post-training on substantial robot data, which can be costly to collect through methods such as teleoperation. Agentic harnesses can adapt around the model, but current self-evolving harnesses use robot trials inefficiently when deciding which code and skill changes to pursue. We introduce EmbodiedEvo, a self-evolving agentic harness that autonomously decides where to explore next and turns the resulting physical interaction into improved code and skills. EmbodiedEvo realizes this through a Fast-Slow Dual-System Architecture, in which competing code and skill hypotheses are maintained in a Hypothesis Graph. Value-of-Information Experiment Selection chooses physical experiments that can distinguish these hypotheses. Their outcomes guide Code-Skill Co-Evolution. The Slow System builds Hierarchical Memory, and Reward-Grounded Memory Learning selects effective memory according to their value for later Fast-System improvement. On RoboCasa365, EmbodiedEvo reaches 77.0% overall success and 71.3% on Composite-Unseen, compared with 40.1% for the best baseline. EmbodiedEvo also reaches 86.8% overall success on LIBERO-Pro. Beyond benchmark performance, EmbodiedEvo transfers zero-shot to real-world robot, achieving 71.3% overall success across multiple challenging tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Agent, Robotics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eHSgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the limited efficiency with which existing embodied agentic harnesses utilize costly physical robot interactions. The authors introduce EmbodiedEvo, a self-evolving agentic harness built around a frozen robot foundation model. The goal is to let the harness actively decide what physical experiment to conduct next and transform the resulting evidence into reusable improvements in executable code, skills, and memory.

The framework contains three main components. Value-of-Information Experiment Selection maintains code and skill hypotheses in a Hypothesis Graph and selects physical experiments according to their expected uncertainty reduction relative to their cost. Code-Skill Co-Evolution explicitly evaluates code, skill, and joint effects through matched trials. Reward-Grounded Memory Learning learns memory-management decisions according to their downstream impact on task performance, exploration uncertainty, and physical-experiment cost.

Overall, the work reframes the relatively heuristic execution-feedback-update loop of prior agentic robot systems as an interaction-efficient continual self-evolution problem.

Critical strengths
  1. The overall system is well integrated, and the Hypothesis Graph provides structured tracking of experimental evidence and code-skill relations.
  2. Reward-Grounded Memory Learning is interesting, as memory operations are learned according to downstream utility rather than fixed heuristics.
  3. The method achieves strong results on RoboCasa365 and LIBERO-Pro and is further evaluated on real robots.
Critical weaknesses
  1. The generation of hypotheses and candidate experiments is under-specified and may depend heavily on LLM proposal quality.
  2. The Value-of-Information Experiment Selection module is mainly compared with Random Exploration; stronger baselines such as uncertainty-only, information-gain-only, greedy success gain, or LLM-guided selection would be more convincing.
  3. The cold-start stability of the PPO-based memory manager and correction of early erroneous memories are not sufficiently analyzed.
  4. In the real-robot experiments, SmolVLA is first fine-tuned with 50 teleoperated demonstrations per task before being frozen for evaluation. This setting demonstrates that EmbodiedEvo can further improve a task-adapted VLA, but it does not fully test one of the key motivations of an agentic harness: improving generalization without updating the underlying VLA. An additional comparison using the pretrained SmolVLA without task-specific fine-tuning would better show whether EmbodiedEvo can provide gains purely through the harness.
Questions for authors

The following points should be addressed.

  1. The generation of hypotheses and candidate experiments is not sufficiently clear. It would be helpful to provide more details on how these hypotheses and experiments are proposed, and to what extent the process depends on the quality of the underlying LLM.

  2. The Value-of-Information Experiment Selection module is mainly compared with Random Exploration. It would be beneficial to include stronger experiment-selection baselines, such as uncertainty-only selection, information-gain-only selection, greedy success-gain selection, or LLM-guided selection, to better isolate the contribution of the proposed VOI formulation.

  3. The Reward-Grounded Memory Learning module uses a PPO-based memory manager, but its behavior during the cold-start stage is not sufficiently discussed. How reliable are the memory decisions before enough downstream rewards have been collected? In addition, if incorrect memories are retained or abstracted at an early stage, how effectively can the system identify and correct them later? Further analysis of the stability and error-correction capability of the memory manager would be helpful.

  4. In the real-robot experiments, SmolVLA is first fine-tuned and then frozen for evaluation. While this setting demonstrates that EmbodiedEvo can further improve a task-adapted VLA, one of the main motivations of an agentic harness is to improve generalization without updating the underlying VLA. It would therefore be valuable to additionally evaluate EmbodiedEvo with the pretrained SmolVLA without task-specific fine-tuning, in order to better demonstrate the generalization gains provided purely by the harness.

5 comments
AuthorsOpenReview ↗

Response to reviewer eHSg (1/3)

Thank you for your exceptionally insightful review! We particularly appreciate your deep understanding of the fundamental challenges facing agentic robotics. Your suggestions highlight four important research directions: evidence-grounded hypothesis generation, principled experiment selection, reliable long-term memory, and generalization beyond the environments where agents learn.

We believe these questions are essential for advancing agentic robotics toward more autonomous systems. Your insights not only strengthen our work but also offer valuable directions for future research in this field. Inspired by your suggestions, we have conducted two additional experiments to further validate our central contributions.

We sincerely appreciate the opportunity to address your questions and present additional evidence. Our detailed responses follow.


W1. Hypothesis and Experiment Generation

Thank you for this insightful question regarding hypothesis generation and LLM dependency.

First, hypothesis generation is grounded in physical execution feedback rather than unconstrained LLM generation. As described in Section 3.2 and Algorithm 1, the Fast System uses execution trajectories, task outcomes, and the existing Hypothesis Graph to propose new code-skill hypotheses or refine existing ones. Candidate experiments are then constructed from these hypotheses under interaction-budget and evaluation constraints. Each experiment explicitly identifies the hypotheses it tests.

Second, our core mechanism does not rely on the LLM to judge hypothesis correctness. Hypotheses are evaluated through matched physical trials. Their confidence is updated using Bayesian posterior estimates, and subsequent experiments are selected through cost-normalized Value of Information. The LLM influences which hypotheses are proposed, while physical evidence determines which hypotheses are supported and which experiments are worth pursuing.

We will expand the implementation details of hypothesis generation, candidate construction, and the associated prompts in the revised manuscript.

W2. Stronger Experiment-Selection Baselines

Thank you for your insightful suggestion to evaluate stronger adaptive exploration strategies. As you suggested, we conducted two additional experiments: Information Gain Only, which maximizes expected information gain without cost normalization, and Thompson Sampling, a standard Bayesian adaptive exploration strategy. All methods use the same interaction budget.

Table R1. Experiment Selection Ablation

Selection StrategySuccess@50 ↑Learning AUC ↑
Random Exploration27.1%0.309
Thompson Sampling25.7%0.324
Information Gain Only61.8%0.556
Full VOI (Ours)72.9%0.635

Full VOI outperforms Information Gain Only by 11.1 percentage points in Success@50, while improving Learning AUC from 0.556 to 0.635. It also outperforms Thompson Sampling by 47.2 percentage points, achieving 1.96× higher Learning AUC.

These results directly isolate the benefit of cost-normalized information gain and demonstrate that our VOI formulation substantially outperforms both uncertainty-reduction-based selection and Bayesian posterior sampling under the same interaction budget.

AuthorsOpenReview ↗

Response to reviewer eHSg (2/3)

W3. Cold-Start Stability and Memory Error Correction

Thank you for your thoughtful observation regarding memory reliability, particularly the importance of cold-start stability and error correction for continually evolving robotic agents.

First, the memory manager does not require a fully trained PPO policy to maintain reliable physical evidence during cold start. As described in Section 3.3 and Appendix C, raw execution records (M1) and evidence-based hypothesis updates (M2) are maintained independently of the memory manager's decisions. The manager operates on memory retrieval and abstraction, with Skip always available when no update is selected. Moreover, PPO training occurs during evolution, and the trained manager is frozen during evaluation.

Second, incorrect memories are not permanently committed. Every M3 abstraction preserves links to its underlying evidence. New physical outcomes continuously update hypothesis confidence, while Merge incorporates new evidence and Forget removes outdated or contradicted memories from active retrieval without deleting their original records. The PPO reward further evaluates memory decisions through subsequent task success, uncertainty reduction, and experiment cost, providing feedback for correcting ineffective memory-management strategies.

To quantitatively evaluate cold-start reliability and error-correction capability, we further analyze memory-decision trajectories recorded during evolution, measuring early-stage contradictions, subsequent memory corrections, and correction latency.

Table R2. Cold-Start Memory Stability and Error Correction

MetricResult
Cold-Start Contradiction Rate ↓15.28%
Memory Correction Rate ↑84.90%
Mean Correction Latency ↓3.71 memory decisions

These results indicate that 84.90% of identified contradictory memories are successfully corrected, requiring an average of only 3.71 memory decisions after contradiction detection. This provides additional evidence that early memory errors can be identified and corrected through subsequent physical feedback rather than permanently propagating through the memory hierarchy.

Finally, Table 5 provides empirical support for reward-grounded memory management. Our full method achieves 71.3% Composite-Unseen success and 0.635 Learning AUC, outperforming PPO with task reward only (65.2%, 0.589) and retrieval-based selection (61.7%, 0.556). These results demonstrate the benefit of learning memory decisions from downstream outcomes rather than relying solely on retrieval similarity or immediate task rewards.

W4. Real-World Generalization with a Frozen VLA

Thank you for highlighting the importance of distinguishing backbone adaptation from harness-level generalization. Your suggestion addresses an important question for agentic robotics: how to improve the capabilities of robot foundation models without requiring repeated parameter updates.

First, we clarify that "frozen VLA" refers to the harness adaptation and inference stages, rather than the absence of any prior fine-tuning. In our real-world experiments, SmolVLA is first adapted to the physical SO-101 setup and target manipulation tasks using 50 demonstrations per task. This establishes the low-level manipulation capabilities required for deployment.

Second, the VLA weights remain completely frozen throughout harness evolution and real-world evaluation. Both SmolVLA and EmbodiedEvo use exactly the same fine-tuned backbone. The only difference is the agentic harness, whose code and skills are evolved entirely in simulation and transferred directly to the physical robot.

Importantly, EmbodiedEvo improves real-world success from 46.0% to 71.3% (+25.3 percentage points) without any additional VLA parameter updates. This controlled comparison directly isolates the generalization gains contributed by the harness rather than additional backbone training.


Thank you again for your valuable insights and constructive suggestions. Your feedback has helped us strengthen our experimental evidence, clarify our underlying mechanisms, and better articulate the broader significance of our work. We believe these additions further demonstrate the potential of evidence-driven, self-evolving agentic harnesses for advancing autonomous robotic learning.

AuthorsOpenReview ↗

Response to reviewer eHSg (3/3)

Dear reviewer eHSg,

We are very grateful for your valuable comments, and we spare no effort to address your concerns. Please let us know if there are any additional concerns. We also hope our responses above address your concerns and merit a better confidence score.

Best, Authors

Reviewer_2qoDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

EmbodiedEvo is a self-evolving harness around a frozen VLA. The Fast System keeps code and skill hypotheses in a graph. It tracks each one with a Beta posterior and picks the next matched trial by expected entropy reduction divided by cost (VOI). A joint gain wijw_{ij} decides whether a code change and a skill change work better together. The Slow System keeps three levels of memory and uses PPO to learn memory operations, with a reward based on success gain, entropy reduction, and trial cost.

Critical strengths
  1. The paper turns the usual “LLM proposes, robot executes” loop into an explicit, cost-aware choice of which experiment to run next (Beta posteriors, a three-outcome prediction model, VOI with stopping rules). This is the real new idea compared with existing harness work.
  2. With the same frozen backbone, the gains on composed tasks are large: +31.2pp on Composite-Unseen, and better than Harness VLA in all 8 LIBERO-Pro cells.
Critical weaknesses
  1. The main results measure per-task adaptation with 200 episodes, and the evaluation setup is unclear and possibly unfair.
  • Appendix E says the reported number is online success over the last 100 episodes during evolution. Section 4.1 says evaluation uses separate, disjoint seeds. These two statements do not match.
  • “Unseen” only means unseen during VLA pretraining. The harness is evolved on the test task itself. No experiment evolves on one task and tests on another, so the claim that improvements are “reusable” and “generalize” is not supported.
  • The paper does not say whether the harness baselines got the same interaction budget.
  1. The core mechanism is not tested on its own.
  • VOI is only compared with random exploration. If Table 4 is RoboCasa365 (its AUC matches Table 5), random exploration reaches 27.1%, far below the frozen backbone alone (57.4%). This baseline is broken, not a fair alternative.
  • There is no comparison with an LLM-driven proposal loop or a standard method such as Thompson sampling.
  • Figure 6 tests what is updated, not whether matched comparisons and wijw_{ij} actually help.
Questions for authors
  1. (W1) Which protocol produced Tables 1–3? What interaction budget did the harness baselines get?
  2. (W1) Can you show results where the harness evolves on some tasks and is tested on tasks it never evolved on?
  3. (W2) Which benchmark is Table 4? Can you compare against an LLM proposal loop and Thompson sampling under the same budget, and ablate the matched comparison / WORKS-WITH filtering?
  4. (W2) Please report standard deviations across seeds.
10 comments
AuthorsOpenReview ↗

Response to reviewer 2qoD (1/5)

We sincerely thank you for your detailed review and constructive suggestions. We appreciate the opportunity to further clarify our contributions and provide additional experimental evidence. We hope our responses below address your concerns and merit a higher score.


W1. Clarification of Evaluation Protocols and Cross-Task Generalization

Thank you for your careful and insightful observations regarding evaluation protocols and generalization.

Q1. Which protocol produced Tables 1–3? What interaction budget did the harness baselines get?

We clarify that each simulation benchmark run consists of two distinct stages within a total budget of 200 episodes: harness evolution and held-out evaluation.

First, the initial 100 episodes are used for harness evolution. During this stage, EmbodiedEvo continuously refines its code, skills, and hypothesis graph through physical interaction.

Second, the final 100 episodes are used exclusively for held-out evaluation, producing the results in Tables 1–2. After evolution, the harness is frozen and evaluated on 10 new evaluation seeds, with 10 trials per seed, strictly disjoint from the evolution seeds. This ensures that the reported success rates reflect generalization to unseen evaluation conditions rather than adaptation to previously encountered episodes.

Third, Table 3 follows a separate real-world evaluation protocol, with 30 trials per task and condition. The simulation-evolved harness is directly transferred to the physical robot without additional real-world evolution.

Regarding baseline fairness, Appendix I confirms that our controlled ablations share identical VLA backbone weights and held-out evaluation settings. All harness baselines follow the same 100-episode held-out evaluation protocol, ensuring a fair comparison. Our controlled ablations additionally use matched interaction budgets. We will explicitly document the adaptation budgets of individual harness baselines and clarify the two-stage protocol in the revised manuscript, including the distinction between evolution and held-out evaluation in Appendix E.

Q2. Can you show results where the harness evolves on some tasks and is tested on tasks it never evolved on?

Yes. Our Section 4.3 already directly evaluates this setting through cross-task, zero-shot sim-to-real transfer.

EmbodiedEvo evolves its code and skills exclusively on simulation tasks. None of the five real-world evaluation tasks are used during harness evolution. The evolved harness is then frozen and directly deployed on these previously unseen tasks, without additional real-world harness evolution.

Importantly, the simulation-evolved code-skill harness remains active throughout real-world evaluation. Both conditions share the same frozen, task-adapted SmolVLA backbone.

As shown in Table 3, EmbodiedEvo improves overall success from 46.0% to 71.3% (+25.3 percentage points) across five unseen real-world manipulation tasks, spanning multi-step manipulation, semantic and arithmetic reasoning, and precision grasping.

These results provide direct empirical evidence of task-disjoint generalization and zero-shot sim-to-real transfer of the evolved harness, rather than adaptation restricted to previously encountered tasks.

AuthorsOpenReview ↗

Response to reviewer 2qoD (2/5)

W2. Effectiveness of the Core Mechanisms and Experimental Controls

We thank you for these insightful suggestions regarding the effectiveness of our core mechanisms and experimental controls.

Q1. Which benchmark is Table 4? Can you compare against an LLM proposal loop and Thompson Sampling under the same budget?

Table 4 evaluates learning efficiency using Success@50 and Learning AUC after 50 physical episodes. Figure 7 separately illustrates a representative LIBERO-Pro SPAT-T learning curve.

First, our Random Exploration baseline is already an LLM-driven proposal-and-revision loop. The LLM iteratively generates and refines code-skill hypotheses based on execution feedback.

Furthermore, as requested, we conducted an additional comparison against Thompson Sampling, a standard Bayesian adaptive exploration strategy, under the same interaction budget.

Table R2. Experiment Selection Ablation

Selection StrategySuccess@50 ↑Learning AUC ↑
LLM-Driven Proposal Loop (Random Selection)27.1%0.309
Thompson Sampling25.7%0.324
Full VOI (Ours)72.9%0.635

Full VOI outperforms Thompson Sampling by 47.2 percentage points in Success@50 and achieves 1.96× higher Learning AUC. These results demonstrate the effectiveness of our hypothesis-guided experiment selection over both random exploration and a standard Bayesian adaptive strategy.

Importantly, the frozen VLA's success rate is not a performance lower bound for an agentic harness. Adding a harness changes the robot's execution behavior and does not guarantee monotonic improvement.

This phenomenon is also observed in existing third-party harness methods: In Table 1, Harness VLA, an independently proposed baseline, achieves 55.8% on Composite-Seen, below its frozen Xiaomi-Robotics-1 backbone (57.1%), despite using identical backbone weights. This demonstrates that adding an agentic harness does not necessarily guarantee monotonic performance improvements.

These results demonstrate that harness-level interventions can reduce task success even when the underlying VLA remains unchanged. Therefore, Random Exploration performing below the frozen backbone does not, by itself, indicate an invalid baseline. Instead, our controlled experiment-selection ablations directly establish the effectiveness of principled adaptation.

Q2. Can you ablate matched comparisons and WORKS-WITH filtering?

Our Code-Skill Co-Evolution is explicitly evidence-driven. As defined in Eq. (7) and Appendix H, code-only, skill-only, and joint execution are evaluated from matched initial states. Their measured success differences determine WORKS-WITH relations in the Hypothesis Graph.

Figure 6 provides controlled experimental evidence for the effectiveness of joint evolution, achieving 71.3% on Composite-Unseen compared with 55.6% for Skill-only Evolution.

These results highlight the benefits of jointly refining code and skills using physical evidence. The matched comparisons quantify the contributions of individual and joint updates, while WORKS-WITH relations preserve empirically supported code-skill interactions.

Q3. Please report standard deviations across seeds.

Thank you for this valuable suggestion. As described in Section 4.1, our evaluation uses 10 held-out evaluation seeds, with 10 trials per seed. We additionally computed the standard deviations across these seeds, as shown below.

Table R3. Evaluation Results Across Random Seeds

BenchmarkMean Success Rate ↑Standard Deviation ↓
RoboCasa36577.0%2.98%
LIBERO-Pro86.8%3.99%

These results quantify performance variability across held-out evaluation seeds. Since a complete benchmark run requires approximately one week, we report evaluation-seed variability separately from variability across independently evolved harnesses.

We will include these statistics and clarify the evaluation protocol in the revised manuscript.


Thank you again for your thoughtful and constructive suggestions. Our additional experiments, controlled comparisons, and statistical analysis provide further evidence for the effectiveness of hypothesis-guided experiment selection, code-skill co-evolution, and cross-task generalization. We believe these results strengthen the empirical foundation of EmbodiedEvo and further demonstrate its potential for efficient continual robot learning.

AuthorsOpenReview ↗

Response to reviewer 2qoD (3/5)

Dear reviewer 2qoD,

We are very grateful for your valuable comments, and we spare no effort to address your concerns. Please let us know if there are any additional concerns. We also hope our responses above address your concerns and merit a better score.

Best, Authors

Reviewer_FbYKOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces EmbodiedEvo, a system for continual robot learning that co-evolves executable code and reusable skills around a frozen robot foundation model. A fast component proposes hypotheses and selects physical or simulated experiments using value of information, while a slower component organizes skills and memory. The authors evaluate the approach on RoboCasa365 and LIBERO-Pro, and transfer the learned harness to a tabletop SO-101 arm. The paper reports improvements over a VLA harness baseline in simulation and on five real-robot tasks.

Critical strengths

1.The paper is generally well organized.

2.The introduction clearly states the authors’ motivation.

Critical weaknesses

1.The paper defines Composite-Unseen as the most stringent RoboCasa365 evaluation, yet EmbodiedEvo reports 71.3% success on Composite-Unseen versus 63.1% on Composite-Seen—an 8.2 percentage-point gap in the counterintuitive direction. While different suites may differ in task difficulty, composition, and object distribution, the paper provides no per-task breakdown, difficulty calibration, or statistical test to explain this reversal.

2.The cost model C(t) accounts only for physical execution time, simulator steps, GPU seconds, and reset cost. It excludes the cost of computing VOI itself, generating candidate experiments, proposing and revising hypotheses, invoking the Qwen3.5-122B-A10B operator, embedding-based retrieval, and PPO training. Computing the VOI in Appendix D requires estimating expected entropy reduction over possible outcomes for each candidate; this can grow combinatorially with the number of candidates and the number of hypothesis nodes tested per experiment. The paper does not report candidate-set size, Hypothesis Graph scale, per-round VOI wall-clock time, or approximation strategies. Consequently, the claimed efficiency is primarily physical-sample efficiency rather than end-to-end computational efficiency, and scalability to larger hypothesis spaces remains unestablished.

3.The ablation study lacks key controls for VOI-based experiment selection. VOI is compared only with random selection, which is a weak baseline. Stronger and more directly comparable controls are missing, including greedy selection of the highest-confidence hypothesis, maximum-uncertainty sampling, information gain without cost normalization, and fixed hypothesis-priority scheduling. Without these controls, the paper does not establish whether the gains stem from the VOI formulation itself, from cost normalization, or simply from any non-random adaptive selection strategy.

Questions for authors

NA

5 comments
AuthorsOpenReview ↗

Response to reviewer FbYK (1/3)

We sincerely thank you for your detailed review and constructive suggestions. We appreciate the opportunity to further clarify our contributions and provide additional experimental evidence. We hope our responses below address your concerns and merit a better score.


W1. Computational Cost and Scalability

We appreciate your concern regarding computational efficiency and scalability.

1. This can grow combinatorially with the number of candidates and the number of hypothesis nodes tested per experiment.

To address this concern, we provide both a complexity analysis and additional quantitative benchmarks demonstrating the computational efficiency of EmbodiedEvo.

Importantly, our VOI computation does not require combinatorial enumeration of joint outcomes. As defined in Eqs. (5)–(6) and Appendix D, additive hypothesis entropy and factorized posterior updates enable exact decomposition of expected information gain into independent per-hypothesis contributions, reducing VOI scoring complexity to O(∑ₜ |Sₜ|). Furthermore, the key point is that we already limit each hypothesis to at most six matched evaluations, as illustrated in Appendix E, yielding only 84 possible posterior count states. This can be precomputed without approximation.

To quantitatively evaluate computational scalability, we implemented a benchmark of the exact VOI formulation under progressively larger Hypothesis Graphs and candidate sets.

Table R1. VOI Computational Scalability Benchmark

Hypothesis NodesCandidate ExperimentsHypotheses per CandidateVOI Time / Round
5010040.0069 ms
10020040.0111 ms
2001,00080.0445 ms
5005,00080.2159 ms
1,00010,00080.4509 ms

These measured results demonstrate the efficiency of our exact VOI scoring. Even with 1,000 hypothesis nodes and 10,000 candidate experiments, each evaluating eight hypotheses, VOI scoring requires only 0.45 ms per round. An additional stress test evaluating 250,000 code-skill pairs completes in approximately 7.8 ms. Together, the exact linear-complexity formulation and numerical benchmarks demonstrate that our VOI scoring scales efficiently to substantially larger hypothesis spaces without requiring approximation, while the framework achieves strong physical-interaction efficiency under its 200-episode budget.

Furthermore, our existing ablation against Random Exploration (Figure 7) demonstrates that EmbodiedEvo reaches sustained 100% rolling success within 48 episodes, whereas Random Exploration fails to achieve the same performance within 100 episodes. With an average end-to-end runtime of 30 minutes per episode, EmbodiedEvo requires only 24 GPU hours, compared with 50 hours for Random Exploration without reaching the same success level, demonstrating substantially greater end-to-end wall-clock efficiency.


W2. Effectiveness of VOI-Based Experiment Selection

We thank you for your constructive suggestion to include stronger experiment-selection baselines.

2. The ablation study lacks key controls for VOI-based experiment selection.

In response, we provide an additional ablation to further demonstrate the effectiveness of our VOI-based selection strategy.

Our VOI formulation jointly considers hypothesis uncertainty, expected information gain, and physical experiment cost, enabling the agent to prioritize experiments that maximize uncertainty reduction per unit cost, rather than relying solely on hypothesis confidence or uncertainty.

Importantly, our existing results demonstrate substantial improvements in exploration efficiency. As shown in Table 4, EmbodiedEvo achieves 72.9% success after 50 episodes, compared with 27.1% for Random Exploration, while improving learning AUC from 0.309 to 0.635. Furthermore, Figure 7 demonstrates that VOI-guided exploration reaches sustained 100% rolling success by episode 48, while random exploration continues to fluctuate through episode 100. These results establish the effectiveness of hypothesis-guided experiment selection under a fixed interaction budget.

Table R2. Ablation of Experiment-Selection Strategies

Selection StrategySuccess@50 ↑Learning AUC ↑
Random Exploration27.1%0.309
Information Gain61.8%0.556
Full VOI (Ours)72.9%0.635

Our additional ablation demonstrates that Full VOI outperforms Information Gain Only by 11.1 percentage points in Success@50 (72.9% vs. 61.8%), while improving Learning AUC from 0.556 to 0.635. These results directly establish the effectiveness of cost-normalized information gain beyond non-random adaptive experiment selection.


AuthorsOpenReview ↗

Response to reviewer FbYK (2/3)

W3. Composite-Seen vs. Composite-Unseen Performance

We thank you for this insightful observation regarding the performance difference between Composite-Seen and Composite-Unseen.

3. The paper provides no per-task breakdown, difficulty calibration, or statistical test to explain this reversal.

Importantly, Composite-Unseen evaluates out-of-distribution compositional generalization rather than absolute task difficulty. In fact, this performance pattern also occurs in the original RoboCasa365 benchmark under a different model and training setting: GR00T N1.5 achieves 42.1% on Composite-Unseen versus 40.6% on Composite-Seen with 100% target-demonstration fine-tuning (Table 2), demonstrating that this reversal is not unique to EmbodiedEvo.

More importantly, our Table 1 demonstrates substantially larger improvements on unseen task compositions, as summarized below.

Table R3. Composite Generalization Performance and Statistical Comparison

MetricComposite-SeenComposite-Unseen
Harness VLA55.8%40.1%
EmbodiedEvo (Ours)63.1%71.3%
Improvement (Ours − Baseline)+7.3 pp+31.2 pp

We further quantify the differential improvement through statistical analysis, as summarized below.

Statistical MeasureResult
Differential Improvement+23.9 pp
Relative Improvement Ratio4.3×
95% Confidence Interval[+8.3, +39.5] pp
p-value0.0026

These results show that EmbodiedEvo achieves a 4.3× larger improvement on Composite-Unseen than on Composite-Seen. The reported 23.9-percentage-point differential gain (95% CI: [+8.3, +39.5] pp; p = 0.0026) provides statistical support for stronger relative improvements on unseen task compositions, highlighting the effectiveness of our method.

AuthorsOpenReview ↗

Response to reviewer FbYK (3/3)

Dear reviewer FbYK,

We are very grateful for your valuable comments, and we spare no effort to address your concerns. Please let us know if there are any additional concerns. We also hope our responses above address your concerns and merit a better score.

Best, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear Area Chair and Reviewers,

We sincerely thank you for your insightful and constructive feedback. EmbodiedEvo addresses a fundamental challenge in agentic robotics: enabling robots to autonomously determine which physical experiment to perform next, rather than relying on inefficient trial-and-error exploration. By introducing Hypothesis-Guided Experiment Selection, Code-Skill Co-Evolution, and Reward-Grounded Memory Learning, EmbodiedEvo transforms limited physical interactions into continual self-improvement and reusable capabilities. Our rebuttal further strengthens these three core contributions with additional experimental evidence.

1. 2.84× Higher Exploration Success: Teaching Robots Where to Explore Next (FbYK W1–W2, 2qoD W2, eHSg W2)

EmbodiedEvo enables robots to autonomously identify the most informative physical experiments, achieving 2.84× higher Success@50 and 1.96× higher Learning AUC than Thompson Sampling. Our method reaches 72.9% success and 0.635 AUC, compared with Thompson Sampling (25.7%, 0.324) and LLM-driven Random Exploration (27.1%, 0.309). It also outperforms Information Gain Only by +11.1 percentage points, validating the effectiveness of cost-aware hypothesis selection. Furthermore, our exact linear-complexity VOI computation requires only 0.45 ms to evaluate 10,000 candidate experiments across 1,000 hypotheses. These additional experiments establish the central innovation of EmbodiedEvo: turning physical exploration from undirected trial-and-error into principled, efficient, and scalable experimental decision-making.

2. 1.28× Higher Success Through Autonomous Code-Skill-Memory Co-Evolution (2qoD W2, eHSg W1 & W3)

EmbodiedEvo transforms physical evidence into continuously improving robot capabilities, achieving 1.28× higher Composite-Unseen success through joint Code-Skill Co-Evolution than Skill-only Evolution (71.3% vs. 55.6%). Unlike conventional harnesses that primarily revise execution strategies, our Fast-Slow Dual-System jointly evolves executable code, reusable skills, and hierarchical memory through hypothesis verification and reward-grounded learning. Reward-Grounded Memory Learning further achieves 71.3% success and 0.635 Learning AUC, outperforming retrieval-based memory (61.7%, 0.556) and task-reward-only PPO (65.2%, 0.589). Our additional analysis demonstrates an 84.90% memory correction rate, confirming that accumulated knowledge can be refined through subsequent physical evidence. Together, these results establish a closed-loop self-evolution mechanism in which each physical experiment informs both future exploration and lasting capability improvement.

3. Zero-to-One Breakthrough: From Task-Specific Evolution to Zero-Shot Real-World Transfer (FbYK W3, 2qoD W1, eHSg W4)

EmbodiedEvo demonstrates a zero-to-one advance from task-specific harness evolution toward transferable robotic self-evolution, achieving 1.55× higher real-world success (71.3% vs. 46.0%) across five unseen physical manipulation tasks. Crucially, the harness evolves exclusively in simulation and transfers directly to real robots without additional real-world harness evolution or VLA parameter updates, using the same frozen, task-adapted backbone. On RoboCasa365, EmbodiedEvo further achieves 71.3% Composite-Unseen success (+31.2 percentage points over Harness VLA), delivering 4.3× greater improvements on unseen versus seen task compositions. Our additional statistical analysis confirms this differential improvement (p = 0.0026; 95% CI: [+8.3, +39.5] pp). These findings demonstrate that physical-experience-driven self-evolution can produce reusable execution capabilities that transfer beyond the tasks and environments where the harness was evolved.

Overall, EmbodiedEvo achieves 77.0% success on RoboCasa365, 86.8% on LIBERO-Pro, and 71.3% on real robots. Our rebuttal further reinforces these results through stronger exploration baselines, computational scalability benchmarks, statistically significant generalization analysis, and controlled mechanism ablations, directly addressing the reviewers' central concerns.

We believe this comprehensive evidence firmly establishes EmbodiedEvo's central contribution: enabling robots to decide where to explore next, learn efficiently from physical interaction, and autonomously evolve transferable capabilities. These findings provide compelling support for the novelty, technical soundness, and significance of our work and merit a positive acceptance decision.

We truly appreciate your thoughtful feedback and consideration!

Best regards,
All Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.