Source-linked AI summary
PRISM: Predictive Representation of Interaction Style and Motion for Social Robot Navigation
Bo-Han Chen, Hiromu Taketsugu, Norimichi Ukita
TL;DR
Robot navigation often represents pedestrians through geometric states, leaving individual interaction tendencies implicit. PRISM passively infers continuous ordinal interaction styles and temporal stability from human trajectories, then supplies them to navigation. In controlled crowd simulations, it reduces collision rates relative to geometry-only navigation while producing small changes in navigation time and path length.
Problem
Robot navigation commonly represents pedestrians mainly through geometric states, while individual yielding or assertive tendencies remain entangled with motion.
Method
PRISM passively encodes human interaction trajectories into a continuous ordinal style representation using Rank-N-Contrast and supplies it with a temporal stability score to the navigation policy.
Results
Across five evaluation regimes, PRISM reduces collision rates; in the training-distribution setting, collisions decrease from 0.19 to 0.12 while success increases from 0.81 to 0.88.
Takeaways & Limitations
Within controlled ORCA-based crowd simulations, passive latent-trait inference can support reinforcement-learning-based social navigation.
Takeaways & Limitations
The evidence is limited to synthetic heterogeneity defined by an ORCA parameter, and the results lack confidence intervals, independent PPO seeds, and closed-loop component ablations.
Abstract
from arXiv · showhide
Humans often observe others before interacting and adjust their behavior accordingly. Robot navigation in crowds, however, often represents pedestrians mainly by observed geometric states, leaving individual differences in interaction tendencies implicit. We propose PRISM (Predictive Representation of Interaction Style and Motion), a framework that infers interaction traits from passive observations of human-human interactions. PRISM encodes human trajectories into a continuous ordinal latent space with a transformer encoder trained by Rank-N-Contrast loss, and pairs each inferred trait with a temporal-stability score supplied to the navigation policy. In randomized crowd simulations, PRISM reduces collision rates over the geometry-only baseline and yields small improvements in navigation-time and path-length metrics. These results suggest the utility of passive latent-trait inference for social navigation in dynamic crowds.
1 Introduction
PRISM addresses latent behavioral heterogeneity by passively inferring interaction styles from human-human interactions and providing continuous style representations to navigation policies. It is designed to personalize navigation without perturbing pedestrians or requiring labels.
- Current navigation methods mainly encode pedestrian positions and velocities, leaving yielding and assertive tendencies entangled with geometric motion.
- PRISM passively observes human-human interactions to infer an agent’s Interaction Style without perturbing that agent.
- The framework learns a continuous ordinal interaction-style representation and supplies it to the navigation policy.
- PRISM encodes relative motion during encounters without probing pedestrians or requesting labels.
- Its contributions include Rank-N-Contrast trait inference, temporal confidence estimates, and evaluation showing improved collision-based safety metrics in heterogeneous crowd simulations.
2 Related Work
Prior work models social dynamics, predicts motion, or personalizes forecasting from trajectory histories, but generally does not expose per-person latent traits to navigation policies. PRISM instead targets a compact interaction-style variable for downstream planning.
- Earlier navigation methods use fixed interaction rules or learned social dynamics from observed trajectories without exposing per-person latent traits to the policy.
- Motion-prediction models capture social interactions for forecasting, whereas PRISM infers an interaction-style variable.
- Personalized prediction methods adapt to individual histories or memory, while PRISM learns a temporally consistent style representation for downstream planning rather than open-loop forecasting.
- The PRISM architecture maps trajectories to Interaction Style Vectors, which are structured by Rank-N-Contrast loss and decoded for stiffness regression and input reconstruction.
3 Proposed Method
PRISM encodes observed interaction histories into an ordinal latent representation and supplies the navigation policy with both the inferred style and its temporal stability. Its training combines reconstruction, regression, and Rank-N-Contrast supervision, while confidence estimation tracks embedding volatility over time.
- 3.1 PRISM: PRISM processes 30-frame position-and-velocity histories with a dual-stage Transformer that attends to each target trajectory and its surrounding context.Histories use a robot-centered local frame, and passive updates rely only on motion observations without probing pedestrians or requesting labels.
- 3.2 Rank-N-Contrast Supervision: Rank-N-Contrast orders samples by stiffness-distance, encouraging embeddings of behaviorally similar agents to remain closer than embeddings of dissimilar agents.The loss uses sorted label distances and a temperature parameter, providing an ordinal training signal rather than requiring arbitrary trait discretization.
- 3.2 Rank-N-Contrast Supervision: The encoder combines reconstruction, Rank-N-Contrast, and regression objectives, but the policy receives the latent projection and confidence score rather than ground-truth stiffness labels.Reconstruction and regression heads support representation learning; their stated loss weights are λRecon = 1.0, λRnC = 1.0, and λReg = 0.1.
- 3.3 Temporal Uncertainty Estimation: Embedding volatility is recursively estimated from successive encoder outputs and converted into a confidence score αt, so unstable style estimates receive lower confidence.The estimator uses β = 0.1, maps volatility with an RBF kernel of width γ = 1.0, and holds the estimate fixed when no neighbor is within 2.5 m.
4 Experiments
PRISM is evaluated in randomized PyBullet crowd simulations and combines ordinal style structure, motion grounding, and stiffness regression before integration into HEIGHT. The full representation improves latent trade-offs and reduces collision rates across evaluation regimes, with modest efficiency gains and smaller benefits under dense obstacles.
- Simulation Environment and Crowd Heterogeneity: PRISM is evaluated in a custom 9m×9m PyBullet simulator with 8–12 static obstacles, Turtlebot dynamics, LiDAR observations, and ORCA pedestrians.The policy uses simulated LiDAR, while PRISM consumes noiseless ground-truth pedestrian positions and simulator-provided identities.
- Training and Evaluation: PRISM augments each HEIGHT neighbor feature with an inferred style vector and confidence score, adding nine values while keeping the remaining architecture unchanged.The encoder is frozen, evaluated every 10 control steps, and excluded from policy-learning gradients.
- Latent Representation: RnC configurations increase Manifold Rank above 0.16, whereas reconstruction-only and regression-only variants remain near zero at 0.019, 0.038, and 0.043.The corresponding t-SNE views show clearer blue-to-red stiffness gradients with RnC supervision.
- Latent Representation: The full model achieves stiffness MAE 0.203, reduces reconstruction ADE from 0.771m to 0.027m, and raises Manifold Rank from 0.043 to 0.183.These comparisons support combining RnC, reconstruction, and regression objectives for a compact navigation state.
- Navigation Performance: 37%: PRISM reduces training-distribution collision rate from 0.19 to 0.12, while success rises from 0.81 to 0.88.Across all five evaluation regimes, PRISM reduces overall collision rate.
- Navigation Performance: Navigation time decreases across all settings and path length decreases in four of five, but gains are smaller under dense obstacles because geometric bottlenecks remain.Aggregate rates and means lack confidence intervals, so the findings are empirical evidence from ORCA simulations rather than statistical-significance claims.
5 Discussion
PRISM treats pedestrians as interactive agents whose inferred interaction styles can inform closed-loop navigation. In ORCA-based simulations, this improves collision safety beyond geometric observations, but evidence remains limited to synthetic behavior and incomplete evaluation.
- PRISM models crowds as dynamic multi-agent worlds and uses passively inferred interaction styles as state variables for closed-loop robot planning.This framing targets settings where robots must reason about how other agents respond during interaction.
- PRISM reduces collision rates across all five ORCA-based evaluation regimes, with only small changes in navigation time and path length.Personal-space violations are penalized independently, so the policy is not explicitly rewarded for exploiting yielding pedestrians.
- Current evidence is limited to synthetic heterogeneity defined by an ORCA parameter, while real behavior is context-dependent and multidimensional.The evaluation also lacks confidence intervals, independent PPO seeds, and closed-loop component ablations.
6 Conclusion
PRISM uses passive interaction-style inference to support behavior-aware robot navigation. In controlled ORCA-based simulations, its trajectory representation is ordered by simulated social stiffness and improves collision-avoidance metrics over geometry-only navigation.
- PRISM demonstrates that passive latent-trait inference can support reinforcement-learning-based social navigation within controlled ORCA-based crowd simulations.Table 2 reports improved collision-avoidance metrics over a geometry-only baseline.
- PRISM organizes trajectory embeddings by a simulated social stiffness parameter and injects the inferred representation into a robot navigation planner.Table 1 and Fig. 3 show a latent representation ordered by interaction style and grounded in observed motion.
- The present results do not solve the sim-to-real problem of obtaining stiffness labels for real humans.