Human-Aware Robot Navigation in Dynamic Crowds using Reinforcement Learning

Socially compliant robot navigation in dense human crowds using reinforcement learning and interaction-aware rewards.

Objective

Build a navigation policy that remains safe, efficient, and socially compliant in dense (specially if group of human’s are also present), dynamic crowds.

My role

  • Designed the RL formulation (state, reward, training protocol)
  • Integrated perception + policy + planning into a usable navigation stack
  • Built evaluation scenarios across multiple crowd densities

Methods

  • First we build a modular reactive based group avodance algorithm named TAGA(A Tangent-Based Group Avoidance Controller)
  • Then moved to RL policy training (sim-to-sim evaluation)
  • Reward shaping for comfort + safety + efficiency
  • Metrics: Group Collision rate(our proposal metric for group navigation performance)

Results (high-level)

  • Summarize best-performing policy settings
  • Key tradeoffs observed (speed vs comfort, etc.)

Media

In website many results are shown. So, please visit the website for see the results.