Stanford's constraint-decomposed RL keeps follower robots close yet safe in dense crowds

Following a specific person through a crowded space is a constant tug-of-war: get too close and you risk collisions; hang back and you lose your target. Stanford researchers—Shiting Gong, Jianpeng Yao, Jinfeng Wang, Marco Pavone, and Jiachen Li—tackle this in a new IROS 2026 paper (arXiv:2608.10056) by rethinking how reinforcement learning encodes the trade-off. Existing RL methods squash competing goals into a single dense reward, making the proximity-safety balance implicit and hard to tune across different crowd densities. The team instead decomposes the task into a sparse task reward and independent cost constraints, each managed through cost thresholds that directly correspond to behavioral meaning—like minimum safe distance or maximum allowed target gap—rather than arbitrary reward weight ratios.

The approach also quantifies prediction uncertainty for pedestrian motion and feeds those estimates into the RL costs, making the robot more cautious when behaviors are unpredictable. The authors ran extensive experiments in both in-distribution and out-of-distribution settings, showing their method achieves a more effective proximity-safety balance than baselines. They then validated it on a real robot in real-world scenarios, confirming the method works beyond simulation. By making the safety-keeping trade-off explicitly tunable, this work gives robotics engineers a practical knob to dial for environments like crowded hospitals, warehouses, or sidewalks—where losing a target or bumping into a bystander are equally unacceptable outcomes.