acceptodds
Withdrawn from ICLR 2027

Learning When to Adapt: Reactivity-Conditioned Residual Adaptation Under Interaction Shift

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pedestrians differ in whether they respond to a mobile robot, and these responses can change during navigation. Policies trained under a single response assumption may therefore be unnecessarily conservative or rely on cooperation that does not occur. We propose reactivity-conditioned residual adaptation (RCRA), which uses explicit estimates of pedestrian reactivity to modulate a learned residual around a frozen conservative navigation policy. A Transformer estimates per-pedestrian reactivity from motion history and robot context, while spatial aggregation determines the residual strength without online gradient updates. Across five ORCA-based interaction regimes, the estimator achieves 84.78% scenario-averaged classification accuracy. In heterogeneous and changing local interactions, replacing a Bayesian reactivity estimator with our learned estimator reduces mean collision rate from 6.56% to 4.85% and mean path length by 3.6%, with navigation components otherwise unchanged. Across all five regimes, RCRA reduces mean navigation time by 16.8% and path length by 16.7% relative to the conservative backbone, while mean success rate decreases from 98.35% to 96.05%. On recorded PeRoI trajectories, our estimator improves balanced accuracy over the Bayesian reactivity estimator by 4.2 percentage points. Controlled robot trials with prescribed reactivity labels validate the residual-modulation mechanism on hardware. These results support explicit reactivity estimation as a useful signal for policy modulation under heterogeneous and changing local interactions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot crowd navigation, human–robot interaction, residual policy adaptation

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_9RytOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RCRA is a robot crowd navigation framework that addresses human-robot interaction distribution shifts caused by varying pedestrian responsiveness (some yield while others ignore the robot). It uses a Transformer module to estimate per-pedestrian reactivity from motion histories, spatially aggregates these predictions into a continuous scaling factor based on distance and bearing, and dynamically modulates a low-rank parameter residual (LoRA) around a frozen, conservative base policy without requiring online gradient updates. Across simulation, real-world trajectory datasets, and hardware trials with a Clearpath Jackal robot, RCRA outperforms traditional Bayesian estimation baselines and achieves a 100% success rate on real-robot crowd navigation tests while significantly reducing navigation time and path length.

Critical strengths

(1) The paper elegantly mitigates human-robot interaction distribution shifts by scaling a parameter residual (Δθ\Delta\theta, parameterized via LoRA) around a frozen, conservative base policy (θ0\theta_0) using an online scalar κt\kappa_t. This architecture is highly consequential because it provides guaranteed safety bounds—falling back gracefully to the conservative baseline in ambiguous or highly non-reactive settings—while requiring no online gradient updates and adding negligible parameter overhead (only 0.14% of the base policy). It achieves scaling efficiency and low computational latency even as the base model scales up, addressing a primary hurdle in real-time robot control.

(2) The authors do not rely solely on passive offline metrics; they actively address the policy-induced covariate shift through DAgger-style iterative training, boosting closed-loop classification accuracy from 77.27% to 84.78%. Furthermore, the paper substantiates its claims across a complete evaluation pipeline: extensive CrowdNav simulation regimes, offline evaluation on the real-world human trajectory dataset PeRoI, and physical deployment on a Clearpath Jackal robot (achieving 20/20 successful navigation trials). This thorough progression from theory to real-world hardware strongly validates the practical utility and robustness of the approach.

Critical weaknesses

(1) The core logic of RCRA relies entirely on the online reactivity estimator to continuously compute the scalar adaptation scale κt\kappa_t. However, evaluation on real-world trajectories (PeRoI dataset) reveals a noticeable performance drop, with balanced classification accuracy falling to 73.0%. In dense or crowded settings, distinguishing whether a pedestrian is yielding to the robot versus simply altering their path to avoid another pedestrian or static obstacle becomes inherently ambiguous. If the estimator misclassifies an aggressive or non-reactive pedestrian as cooperative, the scaling mechanism will shift parameters toward the assertive residual policy (Δθ\Delta\theta), creating significant safety risks and potential collision hazards in tightly coupled social spaces.

(2) While physical experiments on a Clearpath Jackal robot demonstrate proof-of-concept, the testing is constrained to only 4 pedestrians following scripted, pre-determined behavioral roles. Scripted human behaviors fail to capture the high entropy, non-verbal negotiations, occlusions, and sudden trajectory changes characteristic of real-world crowd dynamics. Consequently, the 100% success rate under low-density, controlled conditions does not fully substantiate the paper's claims regarding long-term robustness and safety in unconstrained, high-density public environments (e.g., transit hubs or crowded corridors).

Questions for authors

(1) When the reactivity estimator fails (e.g., misclassifying a non-reactive/aggressive pedestrian as highly reactive in dense or ambiguous contexts), what safety fallbacks or bounding mechanisms prevent κt\kappa_t from scaling up the assertive residual policy Δθ\Delta\theta into a collision trajectory? Specifically, is there an explicit uncertainty threshold or variance measure in bi,tb_{i,t} that forces a fallback to the frozen conservative base policy θ0\theta_0?

(2) How does the Transformer estimator isolate a pedestrian's reaction to the robot from their trajectory adjustments caused by avoiding other pedestrians, group dynamics, or static layout constraints? Was the estimator trained and evaluated under multi-pedestrian interaction dynamics where pedestrian-pedestrian coupling dominates?

(3) In the real-robot Jackal experiments, were trials conducted with unscripted pedestrians who were completely blind to the underlying system rules, or were all 20 trials restricted strictly to the 4 scripted roles? If unscripted trials were run, what was the safety and success rate under natural human walking conditions?

Reviewer_yc4oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RCRA for robot navigation among pedestrians whose responses to the robot differ across individuals and change over time. Specifically, a Transformer estimates pedestrian reactivity from motion history and robot context, which is spatially aggregated to modulate a LoRA residual around a frozen conservative policy without online gradient updates. Evaluation covers 5 simulated interaction regimes, recorded human trajectories, and controlled robot trials. As for the experimental results, RCRA reduces navigation time by 16.8% and path length by 16.7%, while success decreases from 98.35% to 96.05%, relative to the conservative backbone.

Critical strengths
  1. The paper focuses on a practical navigation problem and is well motivated, i.e., in real-world, pedestrians do not respond uniformly to robots. Its evaluation across 5 interaction regimes captures differences between individuals, spatial regions, and changing behavior, providing a useful setting for studying adaptation.
  2. In experimental results, it shows that the learned estimator produces measurable navigation benefits. In Table 3, classification accuracy improves from 72.16% to 84.78% over BDDS. With the other navigation components unchanged, collision rate decreases from 6.56% to 4.85% across the 3 heterogeneous regimes, making the contribution of improved estimation reasonably clear.
  3. The simulation evaluation is substantial, with 1250 test episodes per regime across 5 seeds. The authors also examine estimation under the adaptive policy. 3 DAgger iterations improve accuracy from 77.27% to 84.78%, showing that collecting observations from the deployed policy matters.
  4. The adaptation mechanism offers meaningful efficiency gains. Compared with the conservative backbone, navigation time decreases by 16.8% and path length by 16.7%. In fully reactive crowds, path length falls from 16.19 to 11.01 m while success remains 99.28%. This suggests that the method can exploit pedestrian cooperation without remaining unnecessarily conservative in that setting.
  5. As for the results, the paper evaluates reactivity estimation on recorded human trajectories and demonstrates online navigation on a real robot. Balanced accuracy on PeRoI improves from 68.8% to 73.0%, while the hardware experiments achieve 20 successes in 20 trials. These experiments and accompanying videos support the practical feasibility of residual adaptation, although the hardware trials use prescribed reactivity labels rather than the learned estimator.
Critical weaknesses

My main concerns about the paper are:

  1. Navigation time and path length are measured only on successful episodes. Since the methods succeed on different subsets of trials, the reported reductions of 16.8% and 16.7% may partly reflect differences in episode composition. Comparisons on jointly successful episodes would make the efficiency gains easier to interpret by holding the evaluated routes and interactions consistent.
  2. The hardware experiments do not evaluate the complete learned system. As mentioned in strengths, all 20 trials per policy use prescribed labels for 4 scripted pedestrians. The results support residual modulation, but leave open whether estimation errors and perception noise would undermine its benefits during actual robot navigation. The reliability of the combined estimation and control pipeline therefore remains unverified on hardware.
  3. The real human estimation evidence is relatively limited. PeRoI evaluation uses 208 tracks with a single 70/30 split, excludes 18 attraction tracks, and applies one behavior label throughout each track. Without uncertainty estimates or repeated splits, the reliability of the 4.2 percentage point improvement is difficult to assess. Track labels also cannot directly validate detection of changes within an interaction.
  4. The behavioral advantage of LoRA over simpler action interpolation is small in the reported comparison. With ground truth labels, success is 96.72% versus 96.56% in Mixed and 96.88% versus 96.72% in changing interactions. The computational benefits are clearer than the navigation benefits, particularly since these alternatives are not compared using the learned estimator. Their relative sensitivity to estimation errors remains unclear.
  5. The continuous reactivity scores are used directly for control, but evaluation mainly reports classification accuracy at a threshold of 0.5. The overall accuracy of 84.78% does not establish whether these scores are calibrated. In changing interactions, accuracy falls to 77.1%, making confidence calibration and response delay particularly relevant. An incorrect but confident estimate could activate substantially more residual adaptation than an uncertain prediction.
Questions for authors
  1. Could the authors clarify whether RCRA retains an efficiency advantage at matched collision rates? In changing interactions, collisions increase from 1.12% to 5.52%, so this comparison could materially affect the interpretation of the gains.
  2. Do the authors have existing hardware trials using the learned estimator rather than prescribed reactivity labels? If so, how do estimation errors affect navigation performance? Please try to add experimental results on these end-to-end setting, even if the results are not as perfect as current one, authors should discuss the robustness or the limitations.
  3. Could the authors provide uncertainty estimates for the PeRoI improvement from 68.8% to 73.0% and clarify its consistency across track splits?
  4. Do the authors have existing calibration or response delay measurements for the reactivity estimator? Since continuous scores directly control residual strength, classification accuracy alone may not capture the errors most relevant to navigation.
Reviewer_nvsUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Reactivity-Conditioned Residual Adaptation (RCRA) to address distribution shifts caused by heterogeneous pedestrian responses in human–robot navigation. The method estimates pedestrian reactivity from observed trajectories using a Transformer-based predictor trained with dagger, and dynamically adjusts a frozen navigation policy through a scalar-controlled LoRA residual. Experiments across simulated crowd scenarios demonstrate improvements in navigation efficiency.

Critical strengths
  1. The paper addresses an important challenge in human–robot navigation by explicitly modeling heterogeneous and time-varying pedestrian responses. 2.The proposed method enables lightweight online policy modulation without gradient updates, with computational advantages supported by scalability experiments.
Critical weaknesses
  1. The novelty is limited. Dynamic LoRA scaling and context-conditioned adapter modulation have been explored in prior work
  2. Although the paper explores safety–efficiency trade-offs using ground-truth reactivity, the learned-estimator version exhibits obviously higher collision rates than GenSafeNav in several scenarios.
  3. Although the simulations are closed-loop, pedestrian behaviors are generated by ORCA, with reactivity defined by whether agents account for the robot. It is difficult to simulate the scenario of reactive behavior between human and robots, and quantify improvements of the proposed method.
Questions for authors
  1. Can the authors compare RCRA against a learned context-conditioned gating network that directly predicts the LoRA scaling coefficient from the same observations? What advantages does explicit reactivity supervision provide?
  2. Since dagger uses simulator-provided ORCA reactivity labels, how well does the estimator generalize to unseen or real-world pedestrian interaction policies?

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

Related papers

Discussion (0)

Sign in to comment.