Skip to content

attngraph

Arena wrapper for Intention-Aware AttnGraph (selfAttn_merge_SRNN + GST), the Liu et al. ICRA 2023 crowd navigator. Adapted from Shuijing725/CrowdNav_Prediction_AttnGraph.

Two stacked components:

  • GST predictor (gst_updated/): Gumbel Social Transformer forecasting each visible pedestrian's position 5 steps (1.25 s) ahead. Requires a 5-frame observation window.
  • selfAttn_merge_SRNN policy (rl/networks/): recurrent value/actor network with per-human-node GRUs and human-human spatial attention. Predicted future positions are concatenated into each human's spatial edge feature (12-dim per human).

Run

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

Requires a global plan. Defaults to nav2/navfn.

Files

  • planner.py: bridge wrapper. Maintains a 5-frame pedestrian history and the policy's recurrent hidden state across timesteps, clears both on on_reset.
  • arguments.py, rl/, gst_updated/: vendored upstream code, lightly patched (see ATTRIBUTION.md).
  • planner.yaml: observation manifest.
  • weights.yaml: pulls policy + GST checkpoints from arena-rosnav/attngraph on arena feature planners add attngraph.

License

MIT (inherited from upstream).