rlrvo

Arena wrapper for rl_rvo_nav, a GRU-based actor-critic policy trained with Reciprocal Velocity Obstacle (RVO) collision-time reward shaping. Adapted from hanruihua/rl_rvo_nav.

Run

arena launch robot.mobile:=drl robot.mobile.planner:=rlrvo

Requires a global plan. Defaults to nav2/navfn.

Files

• planner.py: entry point. Builds proprioceptive + per-pedestrian VO cone observations, runs the GRU policy, projects incremental holonomic velocity to [v, omega].
• policy_rnn_ac.py: vendored upstream rnn_ac actor-critic.
• planner.yaml: observation manifest.
• weights.yaml: pulls the pretrained checkpoint + arg file from arena-rosnav/rlrvo on arena feature planners add rlrvo.
• See ATTRIBUTION.md for code provenance.

License

MIT (inherited from upstream).