acceptodds
Withdrawn from ICLR 2027

Probe2Act: Knowing When to Stop Probing Articulated Objects

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A robot can discover how an unfamiliar drawer or hinged door moves by probing it, but every additional interaction consumes budget. We study a practical decision: when does the observed motion justify stopping identification and beginning execution? Probe2Act combines an analytic common-path estimator with a calibrated stopping rule. The estimator retains displacement between probes, so individually weak responses can jointly reveal a local motion direction; the rule stops when the direction's calibrated angular radius meets a declared tolerance, or abstains when the budget is exhausted. Under exchangeability of complete trajectory bundles, the finite-prefix construction bounds the marginal joint event of executing with excessive unsigned direction error. In Panda simulation with supplied contact and link tracking, using 100 calibration and 100 test geometries, Probe2Act reduces clean interaction cost from 12 to 1.87 queries, with 97.5% versus 98.5% task success. Under combined pose noise, it uses 6.78 queries while both methods achieve 98% success. None of its 800 nominal executions has an initial direction error above . Matched comparisons and ablations identify which estimation and uncertainty choices support these savings. Tracking shifts reduce uncertainty-set coverage despite accurate selected directions, identifying a boundary for transfer to real sensing.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Articulated Object Manipulation, Interactive perception, Calibrated stopping, Common-path estimation, Conformal Calibration, Uncertainty-Aware Control

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_8S9qOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates how to efficiently identify the motion of articulated objects, aiming to reduce unnecessary probing while maintaining high manipulation success rates.

Critical strengths

The paper includes comparisons with multiple baselines, ablation studies, and robustness analyses under different noise conditions.

Critical weaknesses
  1. Insufficient literature review and questionable problem novelty. The literature review is notably limited, and the motivation for the proposed research question is not sufficiently justified in light of existing work. For example, the Tac-Man series has already demonstrated the ability to leverage information from probing interactions to directly guide subsequent manipulation actions through closed-loop feedback, without requiring an explicit stopping mechanism. The authors should clarify how the proposed problem differs from, or advances beyond, these existing approaches.

  2. Limited methodological novelty and questionable formulation. The methodological contribution appears limited, and several aspects of the formulation require further justification. In particular, screw theory has a well-established mathematical definition, and screw representations should not be conflated with Lie algebra representations. Moreover, the probing actions are predefined rather than adaptively optimized based on previously acquired information. This limits the method's ability to actively select informative interactions and raises questions about its overall efficiency and practical significance.

  3. Lack of real-world validation. The experimental evaluation is conducted entirely in simulation, without any real-world robotic experiments. The method relies on idealized articulation trajectories with consistent motion patterns, whereas real-world manipulation often involves imperfect motion, compliance, contact uncertainty, and complex dynamics. Consequently, it remains unclear whether the proposed method can reliably estimate articulation motion and achieve similar interaction efficiency in realistic settings.

Questions for authors

See the weakness.

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

This paper proposes Probe2Act, a method that uses geometric estimation and offline calibration to decide when a robot has gathered enough motion information from probing an unfamiliar door or drawer to begin opening it so as to avoid consuming budget too much. The listed contributions are that Probe2Act can preserve useful motion evidence, make stopping explicit, and measure the practical tradeoff. This paper validates their results in Panda simulation on sliding and hinged mechanisms, using 100 calibration and 100 test geometries with supplied contact and link tracing. They show that Probe2Act can reduce clean interaction cost significantly while maintaining similar task success.

Critical strengths

This paper provides a principled stopping criterion for motion identification. Instead of using a fixed probing budget, they use a calibrated directional-uncertainty threshold and efficiently identify the available moving direction of the door without exhausting the budget. The interaction is more efficient under their specific estimation with explicit statistical standard.

Critical weaknesses
  1. The novelty of this paper is not clearly stated. Probe2Act applied prior work ideas to the distinct target, an unsigned continuous motion direction, but this paper only establishes the benefits with its own construction. It remains unclear how much this estimator advances prior approaches. That is, current experiments only show that they can reduce probing interactions wile maintaining high observed success, but does not explicitly show how prior work cannot, or works poorer than Probe2Act.
  2. The scope of the evaluated tasks is too narrow, which degrades the complexity and generalizability of this method. There is not enough evidence that Probe2Act can be generalizes to more complex manipulation settings, as most real-world situations do not assume a fixed, single-degree-of-freedom articulation model, and that assuming fixed grasp makes the claims weaker. If real articulated objects have multiple coupled joints, complex contact conditions, or noisy perception, it is unclear which components of Probe2Act can be applied to these more sophisticated tasks.
Questions for authors
  1. Could the authors clarify the specific methodological advances of Probe2Act over other existing motion-based articulation estimation methods?
  2. To what extent does the proposed method depend on the assumption of a fixed, 1-DOF mechanism? In particular, which components would remain valid for multi-DOF or coupled articulated mechanisms?
  3. How sensitive is Probe2Act to the assumptions of supplied grasp and tracking information and the specific fixed physical parameters (0.15m, etc.)? Are these implementation conveniences, or are they actually necessary for the estimator and the statistical stopping guarantee?
  4. What motivates the selection of a 15 degree angular-error tolerance? Is it a setting that shows empirically higher success rate, or just some estimation benchmark?
Reviewer_AJFPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Probe2Act, a framework for deciding when a robot has gathered sufficient motion evidence to stop probing an articulated object and begin execution. The method combines a common-path geometric estimator, which preserves displacement between probes, with a bootstrap-based angular uncertainty scale and a finite-prefix conformal stopping rule. Under exchangeability of complete trajectory bundles and fixed acquisition and measurement procedures, the calibration yields a marginal bound on the joint event of executing with excessive unsigned direction error.

The method is evaluated in controlled Panda simulation on prismatic and revolute mechanisms with supplied contact and link-pose observations. Compared with a fixed-12-probe common-path estimator, Probe2Act reduces mean queries from 12 to 1.87 under clean sensing and to 6.78 under combined pose noise, while achieving 195/200 versus 197/200 and 196/200 versus 196/200 task successes, respectively. The paper also studies estimator components, alternative calibrated uncertainty scales, and tracking shifts, separating directional accuracy, uncertainty-set coverage, and task success.

Critical strengths
  1. Well-motivated and clearly scoped problem. The paper addresses a practically relevant question in robotic manipulation: when is the available motion evidence sufficient to stop identification and begin execution? By explicitly fixing the probing sequence and separating post-contact identification from upstream perception and downstream control, the formulation permits a focused evaluation of adaptive stopping.
  2. Effective preservation of accumulated motion evidence. The common-path estimator retains between-probe displacement that is discarded by separately centering each probe. The scatter decomposition in Section 4.1 clearly explains this distinction, and the ablation in Table 4 provides compelling evidence of its importance: under combined noise, per-probe centering reduces stopping eligibility from 200/200 to 105/200 trajectories after separate recalibration.
  3. Principled and carefully scoped uncertainty calibration. The finite-prefix construction accounts for data-dependent stopping by calibrating over the acquisition prefixes. The paper states its exchangeability assumptions and correctly limits its guarantee to a marginal joint event involving unsigned directional error, rather than conditional reliability, sensing-shift robustness, or task success. This distinction is important for interpreting the stopping rule.
  4. Informative controlled experiments and transparent reporting. The evaluation includes matched fixed-budget controls, component ablations, comparisons with simpler calibrated scores, and tracking-shift diagnostics. The substantial interaction savings are supported by reported direction-error and task-success outcomes. The authors also disclose cases in which simpler alternatives are competitive, task success is lower, or calibration coverage deteriorates outside its assumptions, rather than presenting uniform superiority.
Critical weaknesses
  1. Incremental methodological novelty. The paper combines established geometric linefitting ideas, residual-based uncertainty estimation, bootstrap resampling, and finite-prefix split-conformal calibration. The authors appropriately acknowledge that the conformal reasoning itself is established. Nevertheless, the central methodological novelty appears to lie primarily in the design and integration of these components for the post-contact stopping problem. The paper would benefit from a sharper explanation of which estimator or stoppingrule design choices constitute new technical insights, beyond the demonstrated usefulness of the combined system.
  2. Dependence of efficiency gains on the predefined probe sequence. The paper intentionally studies adaptive stopping under a fixed 12-probe acquisition bank, rather than adaptive action selection. Within this scope, however, the order and informativeness of the predefined probes may materially affect how early sufficient evidence accumulates. Since the reported query savings are measured for one acquisition sequence, their sensitivity to alternative fixed orderings remains unclear. A small controlled comparison, with appropriate recalibration, would help characterize this dependence without requiring an adaptive probing policy.
  3. Limited explanation of the observed cost–success tradeoff. Although Probe2Act substantially reduces probing interactions, its downstream task success is not uniformly equal to that of the fixed-budget common-path estimator. For example, Table 1 reports 190/200 successes versus 196/200 under 1-mm translation noise. The paper appropriately avoids claiming noninferiority and reports that all 800 nominal Probe2Act executions begin with direction errors below 15◦. Thus, these failures should not be attributed directly to inaccurate initial direction estimates. A more detailed analysis of whether success differences arise from earlier stopping, remaining control-step budget, sign-resolution behavior, or downstream execution would clarify the practical tradeoff.
  4. Limited validation of stopping reliability under changed observation models. The theoretical guarantee explicitly excludes tracking shifts, and the paper appropriately presents these experiments as diagnostics rather than counterexamples to its theory. Nevertheless, Table 5 shows that full-prefix coverage can deteriorate severely under correlated noise and initial-outlier shifts, while selected-prefix coverage and selected-direction accuracy remain much higher. This distinction is informative, but leaves open how reliably the calibrated stopping criterion would behave under a broader range of observation-model changes. The strong bootstrap advantage over the excitation proxy under the tested correlated shift (zero versus 133/200 selected-direction violations in Table 3) is encouraging; further characterization would help establish how specific or general that robustness is.
Questions for authors
  1. Distinct methodological contribution. Beyond the experimentally useful integration of common-path estimation, a bootstrap uncertainty scale, and established finite-prefix conformal calibration, which specific design choices do the authors regard as the main new methodological insights? How do these choices relate to the closest existing geometric estimation and calibrated stopping approaches?
  2. Sensitivity to fixed probe ordering. How do stopping indices, selected-direction accuracy, and query savings change under alternative predefined orderings of the same 12 probes, with each acquisition procedure recalibrated appropriately?
  3. Value of bootstrap relative to simpler scores. Table 3 shows that the calibrated excitation proxy can be more query-efficient under nominal translation noise, whereas the bootstrap scale is markedly more robust under the tested correlated-noise shift. Can the authors further explain the mechanism underlying these different operating points, and whether the added bootstrap computation produces a meaningful downstream task-success advantage in matched simulator executions?
  4. Sources of downstream execution failures. Since nominal selected directions remain within the stated angular tolerance even when task execution fails, what factors account for the success differences between adaptive stopping and fixed-budget estimation? In particular, can the authors distinguish effects of stopping time and remaining execution budget from failures of the fixed downstream controller?
  5. Screw-coordinate scaling sensitivity. How sensitive are the direction estimates, calibrated radii, and stopping decisions to the fixed screw-coordinate scaling parameter ℓ = 0.15m, which Section 3 identifies as untested?

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

Related papers

Discussion (0)

Sign in to comment.