Safe and Fair Multi-Quadruped
Collaborative Box Pushing

Yuhao Li, Williard Joshua Jose, Hao Zhang
Human-Centered Robotics Lab, University of Massachusetts Amherst
ICRA 2026 Workshop on Multi-Agent Robotic Systems
Two failure modes on hardware: robots closing to unsafe inter-robot proximity, and one robot idling while the other pushes.

Two failure modes on hardware when only pushing progress is rewarded. Left: both robots converge on the same contact region and close to unsafe inter-robot spacing. Right: role imbalance — one robot does the pushing while the other stays idle.

Abstract

Multi-quadruped collaborative box pushing requires not only task progress but also safe inter-robot spacing and balanced effort. In our tests, reward formulations that emphasize only task progress lead to unsafe crowding or role imbalance where one robot does most of the pushing. We study this question using a standard MAPPO baseline that augments a task-progress reward with team-level shaping terms for proximity safety and effort fairness. Trained with centralized training and decentralized execution using shared-parameter MAPPO, our policy reaches strong success rates among the evaluated reward variants across easy, normal, and hard orientation scenarios (86%, 68%, 26%) while keeping unsafe proximity low and reducing role imbalance relative to progress-only and fairness-only baselines.

Results

Quantitative comparison of box pushing performance across orientation difficulty bins.

BinMethod Success Rate ↑ Comp. Time (s) ↓ Unsafe Prox. Time % ↓ Single-Robot Time % ↓
Easy
30–60°, 1× diag
P+Proximity54.05.00.047.1
Full-Shaping (ours)86.01.83.444.5
P+Fairness72.01.47.547.6
Progress-only54.01.610.151.8
Normal
90–120°, 2×
P+Proximity32.09.60.151.3
Full-Shaping (ours)68.03.41.125.6
P+Fairness22.03.626.729.3
Progress-only64.04.053.749.4
Hard
150–180°, 4×
P+Proximity4.013.24.355.1
Full-Shaping (ours)26.011.318.726.5
P+Fairness4.012.545.824.4
Progress-only18.08.254.430.1

Why single-objective shaping backfires. Adding only proximity or only fairness shaping degrades success relative to the progress-only baseline in the normal and hard bins. Each single shaping term admits a pathological equilibrium: proximity-only rewards the lazy strategy of disengagement, while fairness-only rewards crowding without spatial awareness. Full shaping keeps both team-level terms active in the return, so reducing the overall loss requires attending to proximity and fairness alongside pushing progress.

Poster

BibTeX

@inproceedings{li2026safefair,
  title     = {Safe and Fair Multi-Quadruped Collaborative Box Pushing},
  author    = {Li, Yuhao and Jose, Williard Joshua and Zhang, Hao},
  booktitle = {ICRA Workshop on Multi-Agent Robotic Systems:
               Real-World Collaboration and Interaction},
  year      = {2026},
  url       = {https://openreview.net/forum?id=HBHwvM839C}
}