---
license: mit
library_name: pytorch
pipeline_tag: robotics
tags:
  - robotics
  - social-navigation
  - human-robot-interaction
  - pedestrian-trajectory-prediction
  - model-predictive-control
---

# PeRoI controller — trained weights

`residual_predictor.pt` — the action-conditioned pedestrian-response predictor behind the **PeRoI
social-navigation controller**. A NeuRoSFM residual (`ŷ = SocialForce + learned_correction`) trained on
the real **PeRoI** robot–human interaction dataset; deployed on a robot as a velocity-grid MPC that
anticipates how each nearby person will react to the robot and plans around them.

**The controller code, real-robot integration guide, ROS node, and sanity check live in the private
GitHub repo → `github.com/elmoghany/peroi-controller`.** This HF repo hosts only the weights.

## Download the weights

```bash
pip install huggingface_hub
hf download elmoghany/peroi-controller residual_predictor.pt --local-dir .
# (private repo — run `hf auth login` first with an account that has access)
```

## Use

```python
from peroi_controller import PeRoIController   # from the GitHub repo
ctrl = PeRoIController("residual_predictor.pt", robot_radius=0.30, v_max=0.6)
vx, vy = ctrl.step(robot_xy, goal_xy, {track_id: (x, y), ...}, dt=loop_dt)
```

Architecture: 3-layer MLP residual on a Social-Force prior; input = pedestrian local-frame features
(past 1 s + goal direction + robot relative state + nearest neighbours + robot-condition one-hot),
output = 8×2 future deltas (2 s @ 0.25 s). ~90 KB, runs in <1 ms/step on CPU.

Provenance + metrics + videos: **https://elmoghany.com/crowd-nav**. License: MIT.
Contact: Mohamed Elmoghany (Cornell).