Learning When to Adapt: Reactivity-Conditioned Residual Adaptation under Human-Robot Distribution Shift

Immanuel R. Santjoko; Benedictus C. G. Cinun; Jing Chen; Bin Hu

cs.ROOct 18, 2026

Learning When to Adapt: Reactivity-Conditioned Residual Adaptation under Human-Robot Distribution Shift

Authors: Immanuel R. Santjoko, Benedictus C. G. Cinun, Jing Chen, Bin Hu

Organizations: Department of Electrical & Computer Engineering, University of Houston · Department of Psychological Sciences, Rice University · Department of Engineering Technology, University of Houston

Abstract

Pedestrians differ in whether they respond to a robot, yet many RL policies assume all are non-reactive to prioritize safety. Policies trained without conditioning on local pedestrians' reactivity may therefore be unnecessarily conservative or rely on cooperation that does not occur. We propose RCRA, an online adaptation framework that modulates a learned residual around a frozen conservative navigation policy according to estimated pedestrian reactivity. A Transformer-based estimator infers latent pedestrian reactivity from motion histories and robot context, providing a more accurate signal for navigation adaptation and improving scenario-averaged simulation accuracy by $12.62$ percentage points over the Bayesian baseline, which assumes known pedestrian goals. These estimates are spatially aggregated into an adaptation scale that gates LoRA modules, enabling cooperative behavior near reactive pedestrians and conservative behavior near non-reactive pedestrians. Among baselines achieving SR above $90\%$ in every scenario, RCRA reduces scenario-averaged path length by $16.7\%$ relative to the conservative backbone and by $7.3\%$ relative to randomized-training baselines. It improves mean SR by $14.88$ percentage points over the reactive-trained policy.

Figures & tables

Figure 1: Overview of RCRA . Offline, a conservative backbone is trained on non-reactive pedestrians and then frozen while LoRA residual modules are learned for reactive-crowd behavior. A reactivity estimator is trained on mixed scenarios. Online, the robot estimates per-human reactivity beliefs from partial observations, aggregates them using spatial relevance weights, and uses the resulting scale $\kappa_{t}$ to gate the LoRA residual to adapt the navigation safety–efficiency operating point.

Scenario	Conditioning Strategy	SR (%) $\uparrow$	CR (%) $\downarrow$	NT (s) $\downarrow$	PL (m) $\downarrow$
Non-reactive	BDDS [ 12 ] (Bayesian)	93.60 $\pm$ 0.63	6.40 $\pm$ 0.63	15.08 $\pm$ 0.34	15.02 $\pm$ 0.33
Non-reactive	Ours (Learned)	95.52 $\pm$ 0.69	4.48 $\pm$ 0.69	15.86 $\pm$ 0.45	15.80 $\pm$ 0.43
Reactive	BDDS [ 12 ] (Bayesian)	99.68 $\pm$ 0.33	0.32 $\pm$ 0.33	11.84 $\pm$ 0.18	11.80 $\pm$ 0.19
Reactive	Ours (Learned)	99.28 $\pm$ 0.33	0.72 $\pm$ 0.33	11.02 $\pm$ 0.12	11.01 $\pm$ 0.12
Mixed	BDDS [ 12 ] (Bayesian)	93.60 $\pm$ 1.44	6.40 $\pm$ 1.44	14.10 $\pm$ 0.46	14.07 $\pm$ 0.39
Mixed	Ours (Learned)	96.08 $\pm$ 1.58	3.92 $\pm$ 1.58	13.68 $\pm$ 0.43	13.64 $\pm$ 0.37

Table 2: Effect of reactivity estimation on navigation. We compare BDDS’s Bayesian estimator with our learned estimator, keeping the backbone, LoRA residual, and spatial weighting fixed and changing only the supplied reactivity estimates.

Scenario	Methods	SR (%) $\uparrow$	CR (%) $\downarrow$	NT (s) $\downarrow$	PL (m) $\downarrow$
Non-reactive	ORCA [ 35 ]	66.40 $\pm$ 3.88	29.20 $\pm$ 2.71	23.68 $\pm$ 0.82	15.03 $\pm$ 0.35
	BRNE [ 21 ]	23.04 $\pm$ 1.56	76.00 $\pm$ 1.17	17.07 $\pm$ 0.54	16.58 $\pm$ 0.53
	CrowdNav++ [ 18 ]	79.20 $\pm$ 2.35	20.80 $\pm$ 2.35	14.42 $\pm$ 0.28	13.53 $\pm$ 0.22
	GenSafeNav (Cons.) [ 38 ]	97.20 $\pm$ 0.40	2.80 $\pm$ 0.40	16.59 $\pm$ 0.47	16.52 $\pm$ 0.47
	GenSafeNav (Naive) [ 38 ]	59.52 $\pm$ 2.47	40.48 $\pm$ 2.47	11.61 $\pm$ 0.20	11.58 $\pm$ 0.21
	GenSafeNav (Rand.) [ 38 ]	91.92 $\pm$ 1.45	7.84 $\pm$ 1.28	15.16 $\pm$ 0.35	15.05 $\pm$ 0.34

Table 3: Interaction-shift benchmark. We compare all methods across non-reactive, reactive, mixed, spatial-cluster, and time-varying scenarios over

1{,}250

test episodes per scenario across five random seeds. RCRA improves the empirical safety–efficiency trade-off.

Figure 3: Controlled hardware validation with prescribed reactivity labels. Top: representative adaptation scale $\kappa_{t}$ and quantitative performance over 30 virtual-pedestrian trials per policy. Bottom: representative demonstrations with instructed physical pedestrians using (a) GenSafeNav (Cons.), (b) GenSafeNav (Naive), and (c) RCRA .

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: DAgger refinement. Reactivity estimation accuracy across aggregation iterations.

Scenario	Setting	SR (%) $\uparrow$	CR (%) $\downarrow$	NT (s) $\downarrow$	PL (m) $\downarrow$
Non-reactive	LoRA ( $\kappa_{t}=0.2$ )	93.84 $\pm$ 0.78	6.16 $\pm$ 0.78	15.01 $\pm$ 0.34	14.94 $\pm$ 0.34
	LoRA ( $\kappa_{t}=0.6$ )	83.12 $\pm$ 1.78	16.88 $\pm$ 1.78	13.25 $\pm$ 0.25	13.23 $\pm$ 0.25
	LoRA ( $\kappa_{t}=1.0$ )	73.12 $\pm$ 3.99	26.88 $\pm$ 3.99	12.35 $\pm$ 0.25	12.33 $\pm$ 0.25
	LoRA (Adaptive)	97.04 $\pm$ 0.36	2.96 $\pm$ 0.36	16.08 $\pm$ 0.41	16.01 $\pm$ 0.44
	Full fine-tuning (Adaptive)	97.04 $\pm$ 0.61	2.96 $\pm$ 0.61	16.01 $\pm$ 0.45	15.93 $\pm$ 0.45
	Action-space (Adaptive)	97.04 $\pm$ 0.46	2.96 $\pm$ 0.46	16.00 $\pm$ 0.47	15.93 $\pm$ 0.46

Appendix

Table 4: Adaptation mechanism ablations. Fixed residual scales expose a safety–efficiency trade-off; adaptive LoRA, full fine-tuning, action-space interpolation, and MPC use ground-truth reactivity to isolate online conditioning. Adaptive-Safe and Adaptive-Efficient bias LoRA toward safety and efficiency, respectively, and are omitted in homogeneous scenarios.

Figure 5: Representative mixed-crowd trajectories, comparing GenSafeNav (Cons.), GenSafeNav (Naive), and RCRA . (a) The conservative policy treats all pedestrians as non-reactive and takes a long detour. (b) The naive policy pursues a shorter route but fails to account for non-reactive pedestrians, leading to collision. (c) RCRA adapts to local robot-directed reactivity, yielding near non-reactive pedestrians while moving more assertively near reactive pedestrians.

Backbone	Method	Train (h)	P95 (ms)	Memory (MB)
2.5M	LoRA	6.30	1.90	47.2
	Gated FT	6.44	2.20	56.6
	Action-interpolation	Shared	3.06	89.7
25M	LoRA	6.80	1.97	287
	Gated FT	7.58	2.23	398
	Action-interpolation	Shared	3.03	510

Appendix

Table 5: Computational scalability. FT denotes full fine-tuning. P95 and memory refer to policy inference latency and peak inference memory. Action interpolation shares the trained endpoints.

Robot Policy Adaptation

Robot policy adaptation changes or steers learned robot behaviors to handle new tasks, environments, hardware, and user preferences.

See all 170 papers

Human-Human & Human-Robot Interaction Transformer (H2INT) for Robot Navigation in Dense and Uncertain Crowds
Sep 4, 2026Ao Shen, Kaixi Chen, Shiwei Liu +2Robot Navigation Social Robot Navigation
School of Automation, Beijing Institute of Technology, Beijing 100081, China · State Key Laboratory of Autonomous Intelligent Unmanned Systems, Beijing 100081, China
Performance-Preserving Online Adaptation in Social Navigation via Diffusion Steering
Sep 21, 2026Haruto Nagahisa, Kohei Matsumoto, Yuki Hyodo +1Robot Policy Adaptation Reinforcement Learning
Graduate School of Information Science and Electrical Engineering, Kyushu University, Fukuoka, Japan · Faculty of Information Science and Electrical Engineering, Kyushu University, Fukuoka, Japan
Proxy Policy Steering
Sep 8, 2026Chuanruo Ning, Tianrui Wang, Wei-Chiu Ma +1Robot Policy Adaptation Robot Foundation Models
Cornell University

Sep 4, 2026J/K move · Enter open · S save

Human-Human & Human-Robot Interaction Transformer (H2INT) for Robot Navigation in Dense and Uncertain Crowds

Ao Shen, Kaixi Chen, Shiwei Liu, Fang Deng+1

School of Automation, Beijing Institute of Technology, Beijing 100081, China · State Key Laboratory of Autonomous Intelligent Unmanned Systems, Beijing 100081, China

Robot Navigation Social Robot Navigation Human-Robot Interaction Transformer-Based RL Pedestrian Behavior Modeling Robust RL

Open paper AskPDF arXiv

Safe robot navigation in dense crowds requires reasoning about pedestrian motion and how it may change in response to a robot. However, many learning-based approaches generate pedestrian motion independently of the robot or assume uniform reciprocity, omitting an important source of interaction uncertainty. This paper presents a Human-Human & Human-Robot Interaction Transformer (H2INT), a reinforcement learning framework that retains robot-conditioned changes in pedestrian motion during policy learning while allowing responsiveness to vary across pedestrians. Responsiveness affects the crowd dynamics when the robot is visible but is not supplied as a policy input; the policy must instead infer its consequences from robot-centered relative positions. A two-stage gated Transformer progressively encodes human-human and human-robot relations, while a recurrent policy captures their temporal evolution. A curriculum gradually reduces pedestrian responsiveness to increase interaction difficulty. Simulation experiments demonstrate improved navigation safety and robustness over representative baselines across response conditions and crowd densities, and show transfer without retraining to structurally distinct crowd-flow layouts. Ablations support the hierarchical relational encoding and gated updates. Real-robot deployment further verifies that the learned policy can operate with sparse observations in a physical environment.

Abstract

Robot Policy Adaptation

Explore similar work

Human-Human & Human-Robot Interaction Transformer (H2INT) for Robot Navigation in Dense and Uncertain Crowds

Performance-Preserving Online Adaptation in Social Navigation via Diffusion Steering

Proxy Policy Steering