acceptodds
Withdrawn from ICLR 2027

C-VOIR: Task Inference for Heterogeneous Robot Teams

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Behavior foundation models (BFMs) provide reusable policies, but an unknown deployment reward requires evidence before a policy can be selected. We study online task inference in heterogeneous robot teams whose members have different embodiments, sensors, actions, and probe costs, and communicate only with neighbors. We propose C-VOIR, a cost-aware value-of-information method that weights reports by sensor noise, age, and prediction consistency. It values physical probes by their expected reduction of uncertainty between competing decision labels minus physical cost. Robots exchange their best positive bids, and a strict-majority lease authorizes at most one probe. The team executes a common policy label only when fresh, consistent opinions indicate that no worthwhile probe remains and the decision gap supports a winner; otherwise, it retains baseline actions and no-op probes. Across three evaluation trials, C-VOIR achieves the highest reported epoch-10 running-average return on four open-sourced benchmarks. We also introduce HeteroArm-Mission, a 600-instance MuJoCo benchmark with two- and three-joint arms, heterogeneous observations, and physical probes. C-VOIR achieves the highest reported non-Oracle task return 0.8810 and cost-adjusted return 0.8750, while reducing mean probe cost by relative to Information Gain. Our code is open-sourced at https://anonymous.4open.science/r/Anonymous_code-77E6.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Behavior Foundation Models, Online Task Inference, Heterogeneous Robot Teams, Value of Information, Multi-Robot Systems

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

This paper proposes C-VOIR, a framework for online task inference in heterogeneous robot teams with local observations and neighbor-only communication. The method combines quality-aware report fusion, decision-relevant probing, distributed probe coordination, and cost-aware stopping. The authors evaluate C-VOIR on three ExORL/DMC domains and introduce HeteroArm-Mission, a MuJoCo benchmark with heterogeneous actions, observations, and physical probes. The results show substantial reductions in probe costs while maintaining competitive task returns.

Critical strengths
  1. The paper addresses an important problem involving heterogeneous sensing, embodiment-specific actions, costly physical probes, and decentralized decision-making. The joint consideration of these challenges is practically relevant.
  2. C-VOIR combines quality-aware report fusion, decision-relevant probing, and distributed coordination into a unified framework. The method is clearly presented, with well-defined protocol assumptions.
  3. On HeteroArm-Mission, C-VOIR substantially reduces the mean probe cost from 3.1627 (Information Gain) to 0.3597 while achieving comparable task returns. The ablation studies help demonstrate the contributions of the proposed components.
Critical weaknesses
  1. The proposed decision rule does not directly optimize the team utility defined in Eq. (22). Requiring agreement among locally optimal labels may prevent execution even when a unique team-optimal common label exists.
  2. The data-dependent weighting in Eqs. (4–5) may invalidate the standard Bayesian interpretation of the posterior. The paper does not sufficiently justify the confidence bounds used for action feasibility and stopping decisions.
  3. The NetVOI criterion in Eq. (6) focuses on local uncertainty reduction rather than team-level utility. The relationship between the selected probes and improvements in overall team decision quality remains unclear.
  4. The performance gains over strong baselines are marginal, particularly on HeteroArm-Mission and Cheetah. Only three evaluation trials are reported, and comparisons under matched probing budgets are missing.
Questions for authors
  1. Under what conditions does the SameLabel requirement in Eq. (8) guarantee the team-optimal label defined in Eq. (22)? Can the authors provide a proof?
  2. How does the data-dependent report weighting in Eqs. (4)–(5) affect the validity of the posterior confidence bounds? Can the authors provide theoretical or empirical justification?
  3. Can the authors report team-level regret relative to the optimal common label, along with the common-label feasibility, label-conflict, and fallback rates defined in Appendix F.4?
  4. Can the authors provide paired statistical comparisons and evaluate the performance–cost trade-off against baselines under matched probing budgets?
Reviewer_4reNOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces C-VOIR, a method for online task inference with heterogeneous robot teams using behavior foundation models (BFMs). C-VOIR integrates report fusion, probe selection, and a cost-aware team stopping for robot teams. The experiments on three locomotion tasks and a newly introduced HeteroArm-Mission to show the improved performance.

Critical strengths
  • C-VOIR shows good performance-cost trade-off compared with Information Gain and DecBE-LUCB.
Critical weaknesses
  • One of the main experiments evaluated on DMC locomotion tasks (Walker, Cheetah, Quadruped). These are single-agent locomotion benchmarks, not the heterogeneous robot team scenarios targeted in this paper.
  • Several key components, definition, and information are not included in the paper, including the details about the HeteroArm-Mission benchmark, its statistics, the definition of reports, probes, decision labels, bids, and leases. These are used without concrete grounding or formally defined which make the paper very hard to follow.
Questions for authors
  • What specific aspect of the heterogeneous team is tested in the DMC locomotion tasks (Walker, Cheetah, Quadruped)?
  • Could you provide the details about HeteroArm-Mission, including its task formulation, benchmark statistics, what a probe, report, and decision label are in the two- and three-joint arms.
Reviewer_kJePOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes C-VOIR for online task inference in heterogeneous robot teams when the deployment reward is unknown. The method combines quality-aware observation fusion, decision-relevant probing with a physical cost penalty, and distributed agreement and stopping rules. Its main idea is to reduce unnecessary probing while collecting sufficient evidence to select appropriate policies from a behavior library. The authors evaluate the method on three DMC domains and introduce HeteroArm-Mission for heterogeneous robot evaluation. The reported results show competitive task returns and substantially reduced probe costs, with additional controlled experiments examining evidence fusion, probe selection, and communication.

Critical strengths
  1. This paper addresses a meaningful problem: how heterogeneous robot teams can collect sufficient information for task inference while avoiding unnecessary probing. Combining observation-quality weighting, decision-relevant probe selection, and distributed coordination provides a useful framework for studying this problem.

  2. Table 2 shows substantially lower probe costs than the probing baselines while maintaining comparable task returns. The ablation in Table 9 further shows that focusing on decision-relevant information reduces probing without sacrificing task return, while the cost penalty provides additional savings. These results support the potential of the approach to improve the performance cost trade-off.

Critical weaknesses
  1. Firstly, the comparisons do not sufficiently isolate the source of the probe-cost advantage. Table 9 provides useful evidence that decision-relevant probing reduces unnecessary information collection. However, Information Gain performs approximately six probes, whereas C-VOIR performs only 0.884 on average, so much of the cost reduction accompanies a reduction in probe count. It remains unclear whether similar benefits could be obtained by adding a simple stopping rule or cost penalty to the baselines. Comparisons under matched probe-cost budgets, together with performance cost curves, would help establish whether C-VOIR uses limited probing resources more effectively. Several methods in Table 2 do not perform physical probes at all; these comparisons help assess whether additional probing is worthwhile, but do not establish an advantage in reducing probe cost. Since the performance–cost trade-off is central to the paper, stronger controlled comparisons are needed to substantiate this contribution.

  2. Secondly, the evaluation does not clearly establish how the reported gains translate into effective team execution. On HeteroArm-Mission, task return measures the relative value of selected actions within a fixed candidate library, and team ε-success is defined by a normalized regret threshold. These metrics are relevant to policy selection, but do not directly show whether the robots complete the intended physical task or successfully execute a common team decision. More intuitive measures, such as task completion under a fixed time or energy budget, would help demonstrate the practical benefit. In addition, Appendix F.4 defines common-label feasibility, label-conflict, and fallback/no-decision rates, but their results are not reported. Without these statistics, it might be difficult to assess how often the coordination mechanism authorizes a supported common decision and how often the system falls back to baseline actions. Reporting these outcomes, alongside a representative probing-to-execution example, would make the contribution of the team coordination mechanism clearer.

Questions for authors
  1. Could the authors clarify the stopping rules used by the probing baselines? In particular, does Information Gain in Table 9 continue probing throughout all six rounds, or can it stop early based on a value or cost criterion? This would help clarify how much of the cost advantage comes from selecting more useful probes versus stopping earlier.

  2. On HeteroArm-Mission, what proportion of instances reaches an authorized common decision by the final round, and what proportion falls back to baseline actions? Are these fallback outcomes included in the reported task return and ε-success? Providing the relevant statistics defined in Appendix F.4, if available from the existing evaluations, would help clarify the practical contribution of team coordination.

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

Related papers

Discussion (0)

Sign in to comment.