Belief Dynamics for Detecting Behavioral Shifts in Safe Collaborative Manipulation

Devashri Naik, Divake Kumar, Nastaran Darabi, Amit Ranjan TrivediarXiv 2026 · DOI

TL;DR

This paper introduces UA-TOM, a novel belief-tracking module that enables robots to reliably detect when a collaborating agent changes its strategy mid-task. This is crucial for safety in shared workspaces, as it significantly reduces collisions and improves adaptation compared to existing methods, even outperforming an "oracle" with perfect detection timing.

What Problem Does This Paper Address?

Robots are increasingly working alongside humans and other agents in shared spaces. A critical challenge arises when these collaborators unexpectedly change their behavior or "strategy" (regime) during a task. If the robot continues to operate under its old assumptions about the collaborator's behavior, it can lead to unsafe situations, increased collision risks, and degraded task performance. Detecting these mid-episode behavioral shifts in real-time is essential for safe and effective human-robot collaboration, but current methods often struggle with reliability, especially under strict timing constraints.

Methods

The core contribution is UA-TOM, a lightweight belief-tracking module designed to augment existing "frozen" vision-language-action (VLA) robot control backbones. UA-TOM doesn't retrain the entire robot policy; instead, it adds a specialized layer that focuses on inferring the collaborator's behavior. It achieves this through a combination of:

  • Selective State-Space Model (SSM): This component maintains a persistent belief state about the collaborator's behavior, updating it over time. Crucially, the dynamics of this update are learned to be sensitive to changes.
  • Causal Attention: This helps re-anchor the belief state to recent observations, particularly useful for recovering from sparse information or in discrete coordination tasks.
  • Prediction-Error Signals: This pathway directly signals when the robot's prediction of the collaborator's actions deviates significantly from reality, acting as a strong indicator of a behavioral shift.
  • Contrastive Memory: This component helps maintain categorical cues about different collaborator types, aiding in distinguishing between them.

These signals are fused in a gated integration layer to produce estimates of the collaborator's type and the probability of a regime switch.

Datasets

  • ManiSkill: Used for training and evaluating the primary manipulation tasks. This environment involves two robotic arms in a shared workspace.
  • Overcooked: Used for cross-domain evaluation, demonstrating the generalizability of UA-TOM's architectural components. This is a grid-world game simulating collaborative cooking.

The paper also implicitly uses a "scripted" set of collaborator behaviors (Helper, Competitor, Blocker, Passive) within these environments to simulate regime switches.

Key Insights and Ideas

  • Reliability over Mean Performance: Standard evaluation metrics (like average detection rate) can mask critical failures. The paper emphasizes the importance of worst-case reliability, especially in safety-critical applications.
  • Timing is Everything: Detecting a switch is only half the battle. The speed of detection, particularly within tight operational tolerances (like ±3 steps, equivalent to 150ms in a 50ms control loop), is what truly matters for safety.
  • Smoothness is Safer: UA-TOM's gradual belief revision, driven by its SSM, leads to smoother adaptation post-switch, resulting in lower "close-range time" (time spent near the danger zone) than even an oracle that has perfect detection but might switch strategies abruptly.
  • Domain-Specific Contributions: The effectiveness of different UA-TOM components (e.g., prediction error vs. causal attention) varies depending on whether the task is continuous manipulation or discrete coordination, suggesting a need for modularity in future designs.

Key Results

  • Collision Reduction: Enabling regime-switch detection reduced post-switch collisions by 52% compared to a no-detection baseline.
  • Detection Reliability Tiers: At a strict ±3 step tolerance, methods fall into three tiers: fast detectors (>82% detection), agent-modeling methods (58-62%), and basic sequence models (<34%). Standard ±5 tolerance masks these differences.
  • UA-TOM Performance: UA-TOM achieved the highest detection rate among unassisted methods (85.7% at ±3 steps) and the lowest close-range time (4.8 steps), even outperforming an Oracle (5.3 steps).
  • Internal Dynamics: The SSM's hidden-state update magnitude spikes 17x at regime switches, driven by prediction errors, and decays over about 10 timesteps, explaining UA-TOM's sensitivity and smooth adaptation.
  • Inference Overhead: UA-TOM adds a minimal 7.41 ms inference overhead per forward pass, representing 14.8% of a 50 ms control budget.

Limitations and Future Work

  • Scripted Behaviors: The current evaluation uses scripted collaborator behaviors. Future work should extend to more realistic, continuous human behavior and vision-based inputs.
  • Discrete Types: The system is designed for discrete collaborator types. Adapting to continuous variations in behavior would be a valuable next step.
  • Limited Ablation: The ManiSkill ablation study was single-seed, providing directional evidence rather than definitive proof.
  • Focus on Detection: While UA-TOM enables safer adaptation, the paper primarily focuses on the detection aspect. Further research could explore how to best leverage these detected shifts for more sophisticated coordination policies.

Summary updated

BibTeX
@misc{naik2026belief,
  author = {Devashri Naik and Divake Kumar and Nastaran Darabi and Amit Ranjan Trivedi},
  title = {{Belief Dynamics for Detecting Behavioral Shifts in Safe Collaborative Manipulation}},
  year = {2026},
  eprint = {2604.04967},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.04967},
}

Loading related content...

Figure