acceptodds
Under review as a conference paper at ICLR 2027

Runtime Recovery for Frozen Generative Robot Policies under Long-Tailed Distributions

Abstract

Generative behavior cloning can produce diverse action distributions and represent complex, multimodal robot behaviors. However, robot execution is also governed by physical dynamics, including contact, friction, and object motion. Small action errors can be amplified during execution, leading to substantial deviations in the robot state, especially when the robot enters long-tail states that are weakly covered by the training data. A common approach is to sample multiple candidate actions from the original policy and select a more suitable one for execution. However, if the policy itself does not cover successful actions at the current state, a limited increase in the number of samples may still fail to identify an effective action. To address this issue, we propose TailGuard, a runtime recovery method for generative policies in long-tail states. TailGuard keeps the original policy frozen and uses previously successful rollouts to characterize task-relevant successful regions. When execution deviates from these regions, an external controller, together with a learned short-horizon dynamics model, corrects the robot behavior and guides it back toward states from which successful execution is feasible. Control is then handed back to the original policy once it can reliably continue the task while maintaining action continuity. Across four robotic manipulation benchmarks, TailGuard requires neither additional demonstrations nor failure data, improving the average closed-loop success rate from 49.1% to 70.4% and outperforming retrieval- and residual-based recovery as well as recent recovery methods for diffusion policies.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Generative Robot Policies, Runtime Recovery, Long-Tailed Distributions, Failure Recovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Vxp8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed TailGuard, a runtime recovery framework for frozen generative robot policies under long-tailed distributions. TailGuard builds a task-conditioned success manifold from previously successful rollouts. During execution, it detects deviations from this manifold and uses a learned short-horizon dynamics model to optimize recovery actions. It also introduces a mechanism to determine when control should return to the original policy. The method is evaluated on four simulated manipulation settings using Diffusion Policy and SmolVLA. TailGuard improves the average success rate from 49.1% to 70.4% and achieves competitive results against several existing recovery methods. The paper also provides ablation studies and real-robot demonstrations.

Critical strengths
  1. The paper has a clear motivation and proposes a practical recovery framework. TailGuard provides a reasonable solution by combining success-manifold estimation, dynamics-guided action correction, and a compatibility-based handoff mechanism. I particularly appreciate that the method explicitly considers when to return control to the original policy, rather than focusing only on generating recovery actions.
  2. The experimental results are promising. TailGuard improves the average success rate from 49.1% to 70.4% across four simulated settings. It also performs competitively against recent methods such as REACH, DynaGuide, and LPB. In addition, the ablation studies on the handoff mechanism, feature representations, and recovery parameters provide useful evidence for the proposed design.
Critical weaknesses
  1. The novelty of the method is a bit limited. I think the method mainly builds on existing ideas, such as dynamics-guided policy correction and recovery toward successful states, which have been explored in DynaGuide and LPB. TailGuard mainly combines these ideas with success-manifold estimation and a handoff mechanism. Also, the Retrieval and Residual baselines use different execution and handoff settings, making it difficult to isolate the contribution of the proposed recovery method.
  2. The learned dynamics model is trained on demonstration data, but recovery is mainly performed in long-tail states outside the training distribution. The paper does not sufficiently evaluate whether the model can provide accurate predictions and useful gradients in these states.
  3. Most experiments focus on initial-state perturbations, while real-world recovery is demonstrated only through single-trial examples. This makes it difficult to assess the method’s reliability under more general execution failures.
Questions for authors
  1. Could the authors provide more details on how REACH, DynaGuide, and LPB were adapted to the benchmarks? Also, could the authors compare TailGuard with simpler recovery baselines under the same execution horizon and handoff settings?
  2. How accurate is the learned dynamics model in long-tail states where recovery is triggered? Could the authors report prediction errors across different tail regions and examine whether the predicted improvements from action optimization match the actual outcomes?
  3. Most quantitative experiments focus on initial-state perturbations. Could the authors also evaluate recovery under execution-time disturbances, such as object displacement or recoverable contact failures?
Reviewer_W1SBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper shows how generative policies that fail in unexpected ways not covered by the initial training data distribution can be recovered through an external controller and a learned model. There is significant evaluations on relevant simulation benchmarks and comparison against SOTA methods. Experimental results are limited.

Critical strengths
  • Addresses a fundamental problem of generative policies: what to do if something goes wrong and the data on which the policy is trained does not include such a failure case.
  • Clear presentation and good simulation evaluation on several benchmarks and against SOTA methods, including a detailed appendix
Critical weaknesses
  • Statement in abstract is misleading as the SOTA methods achieve around 68% and this new method 70%; the difference is much smaller than the stated 20% improvement
  • Real-world experiments are very limited: More experiments with more variations and comparisons against the SOTA methods would help to convince me that the method really outperforms SOTA.
  • A further potential contribution would be to show how the method deals with unexpected obstacles that don't appear in the training data, and compare, for example, against the paper: Diffusion Predictive Control with Constraints, L4DC 2025, which considers (new) collision constraints in the denoising process and not as an afterthought.
  • Discuss the connection to the paper: From demonstrations to safe deployment: Path-consistent safety filtering for diffusion policies, ICRA 2026, which can guarantee adaptation to new scenarios and obstacles without altering the distribution of the data the policy was learned from, thereby preemptively avoiding failures.
Questions for authors
  • Where does the learned short-horizon model come from, and how accurate is it? And why must it be learned? Could you not simply do traditional model-based robot motion planning/control?
  • How do you deal with hard constraints such as unexpected obstacles in the way?
Reviewer_49GZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is about with a frozen generative robot policy, how to use a runtime recovery method for generative policies in long-tail states. TailGuard keeps the original policy frozen and uses previously successful rollouts to characterize a task-conditioned success manifold. Therefore when there is deviation from the successful manifold, an external controller would be used to correct the action back to the success manifold. Several experiments show that this can be an effective way to improve the frozen generative policy.

Critical strengths
  1. This paper provides an interesting idea to characterize a successful conditioned manifold. And doing inference time correction for frozen policy with this manifold.
  2. It does useful ablation about the manifold metrics and visual features.
Critical weaknesses
  1. Even though the paper claims the proposed method outperforms retrieval- and residual-based recovery, I am concerned whether it's a fair comparison. Like whether the receding-horizon execution and other factors are consistent and enable a fair comparison. Please illustrate more on this.
  2. The paper proposes that TailGuard doesn't require additional demonstrations or failure data. But it seems training a dynamics model is required? It seems the paper overstates the function of TailGuard.
  3. The dynamics predictor seems to be an important factor in the whole pipeline. There should be more studies about the relationship between effectiveness of the dynamics predictor and the success rate. And what about replacing the dynamics predictor with more powerful world modeling?
  4. The realworld experiment is too limited, and with too little number of trials.
Questions for authors

Please address the questions written in weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.