acceptodds
Under review as a conference paper at ICLR 2027

Stage-Verifiable Latent Chain-of-Thought with Local Repair for Long-Horizon Robotic Manipulation

Abstract

Long-horizon robotic manipulation requires intermediate reasoning to remain consistent with observed task progress. Continuous latent plans, however, often lack stage semantics that can be checked during execution. We present a vision-language-action framework that aligns latent segments with demonstration-derived object relations and subgoal completion, verifies their expected outcomes against observations, and repairs the affected suffix after execution deviations. The framework retains a verified prefix while its still-required conditions remain valid and conditions regeneration on the current scene and task history. All model calls, including verification and repair, are included in a task-level inference budget. The method achieves measured mean success rates of 98.0% on LIBERO overall and 96.8% on LIBERO-Long, and 58.0%/23.0% under clean/randomized conditions on 50 RoboTwin 2.0 tasks. On LIBERO-RECOVER, enabling repair increases macro-averaged recovery success from 22.8% to 26.0%. Relative to full replanning, local repair reduces cumulative inference time from 18.6 to 14.9 seconds per episode, a 19.9% reduction, with recovery success of 26.0% versus 26.4%. These results characterize the recovery–computation trade-off under the evaluated conditions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robotic Manipulation, Long-Horizon Planning, Vision-Language-Action Models, Latent Chain-of-Thought, Execution Verification, Failure Recovery

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xERoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes a stage-verifiable implicit thinking chain framework for long-term robot operations. It achieves observation-driven stage verification by aligning the implicit planning segments with the completion states of sub-goals. When there is an execution deviation, it performs local suffix repair based on task dependencies, achieving a balance between fault recovery capability and computational cost to a certain extent.

Critical strengths
  1. The proposed method introduces verifiable stage semantics into implicit planning, enabling the verification of sub-goals based on observations without the need for text planning.
  2. The proposed method has been experimentally verified on multiple benchmarks.
Critical weaknesses
  1. The proposed method is overly complex and does not bring significant advantages.
  2. Stage alignment is achieved by relying on the annotations obtained from simulation; all experiments are conducted solely using simulation and have not been verified on real robots.
  3. The results are based on a specific reasoning budget; the trade-off relationship will vary under different hardware and budget conditions; the small performance differences do not have statistical significance.
Questions for authors

See the above weaknesses.

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

This paper introduces a stage-verifiable latent chain-of-thought framework for long-horizon robotic manipulation. Continuous latent planning segments are aligned with demonstration-derived object relations and subgoal-completion signals, so that intermediate plans can be checked against observations at execution time. An observation-based estimator evaluates the executed state, and a dependency-aware repair procedure retains valid latent prefixes while regenerating only the affected suffix after persistent deviations.

The framework builds on OpenVLA with continuous action chunks and is evaluated on LIBERO, RoboTwin 2.0, and LIBERO-RECOVER, reaching 98.0% average success on LIBERO, 96.8% on LIBERO-Long, and 26.0% macro-averaged recovery success. Relative to full replanning, local repair cuts cumulative inference time by 19.9% with recovery success of 26.0% versus 26.4%. The main contribution lies in unifying stage-aligned latent representations, observation-based verification, and selective plan reconstruction within a single VLA framework.

Critical strengths
  • A meaningful verification interface via stage-aligned latents. By supervising latent stages with object relations and subgoal completions, and separating predicted outcomes from observation-based estimation, the paper avoids treating a plausible latent prediction as proof of physical success. Table 4 supports this: removing alignment drops verification F1 from 92.4% to 81.6% and recovery from 26.0% to 22.1%.

  • Controlled ablations and full cost accounting. The experiments isolate state supervision, stage alignment, repair, and repair scope, and the cost analysis includes verification and repair overhead rather than vector counts alone. Table 5 shows local repair halves p95 latency (820->430 ms) and saves 19.9% inference time versus full replanning at nearly equal recovery, providing useful evidence on the recovery–computation trade-off.

Critical weaknesses
  • Dependency-aware repair is underspecified. The mechanism is central to the paper, yet how prerequisite dependencies between stages are obtained (manual, annotation-derived, or learned) and how AtA_t is computed online are not clearly explained, nor are the exact decision rules and thresholds. This limits reproducibility and makes it hard to judge how much of the recovery behavior stems from learned latent reasoning rather than externally supplied task structure.

  • Evidence of advantage over simpler strategies is weak. The benefits attributable to the repair mechanism are small — +3.2 points recovery over repair-disabled, +1.4 over current-stage-only repair, and slightly below full replanning (26.0% vs 26.4%) — reported as three-seed means without uncertainty estimates. Combined with low absolute performance in hard settings (L4: 9.0%; RoboTwin randomized: 23.0%) and no real-robot evaluation, the results support a trade-off point under the tested conditions but not a convincing general advantage.

Questions for authors
  1. Dependency construction and repair localization. How are prerequisite relations between stages obtained — manually defined, extracted from demonstrations, or predicted? Please give a concrete algorithm for building AtA_t, including how uncertainty in prerequisite validity is handled.
  2. Stage prediction and verification. How are stage boundaries and termination predicted at inference, and under what training objective? What are the verification thresholds, confidence-weighting rules, and persistent-deviation criteria, and how sensitive are results to them?
  3. Statistical reliability and online verification accuracy. Please report seed-wise deviations or confidence intervals for the main ablations, plus online failure-detection precision/recall, false-trigger rates, and repair-localization accuracy, rather than only diagnostic-set F1.
  4. Computation budget and recovery efficiency. What is the exact budget BB? Please provide recovery-success-versus-cost curves across budgets, along with episode lengths and triggered-repair counts, to separate prefix-reuse savings from termination effects.
  5. Role of recovery supervision. Table 6 shows recovery data lifting success from 26.0% to 31.0%; how does the model generalize from demonstration-derived suffix training to off-demonstration failure states, and where does correct localization still fail at the action level?
Reviewer_T2M5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper propose a VLA policy (built on OpenVLA-OFT) whose plan is a sequence of latent stages. Each stage is supervised to decode into object-relation and subgoal predicates from demonstrations, and a separate observation estimator predicts the same predicates from the scene. At test time the two are compared to confirm progress or trigger a repair, which keeps the still-valid prefix and regenerates the rest of the plan. Results are reported on LIBERO, RoboTwin 2.0 and LIBERO-RECOVER.

Critical strengths
  • The motivation is good. Latent reasoning in VLAs is usually judged only by final success, and tying latent stages to checkable predicates so the plan can be verified and partially reused during execution is a good direction.

  • The repair-scope comparison (Table 5) counts all model calls and compares local repair, current-stage repair and full replanning under the same setup. I found this the most useful part of the paper.

Critical weaknesses
  • The paper is quite dense and leaves out a number of implementation details (predicate set, trigger thresholds, budget value), which made it hard to follow.

  • The gains are small. LIBERO is close to saturated, repair adds about 3 points of recovery, and local repair vs. full replanning is 26.0 vs 26.4. Only three-seed means are given, so I can't tell whether these differences are real. Adding recovery data to training (Table 6) helps more than the repair mechanism itself, and the claimed time savings may partly come from episodes ending differently.

  • The key baseline is missing. The verifier is basically a predicate classifier trained on simulator labels. With the same labels, an explicit stage/task-graph monitor that retries from the violated precondition is the obvious alternative, and the authors say they do not compare against it. Without it, it is unclear what the latent formulation adds. It is also not clear how the "still-required conditions" for dependency-aware repair are obtained at test time.

Questions for authors
  • Can you report variance over seeds for Tables 4 and 5?

  • How does local repair compare with an explicit predicate/task-graph monitor using the same estimator?

  • Does the dependency-aware repair help mainly at the harder recovery levels, where earlier prerequisites are actually broken?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.