---

Spot RL Inference

Real-time deployment runtime for running reinforcement learning policies on the Boston Dynamics Spot quadruped. It takes ONNX policies exported from spot_rl_train and runs them on the real robot through the Spot SDK.

The runtime streams robot state from Spot, builds the same observation vector used during training, runs the policy, and streams joint position commands back to the robot.

Installation

This project uses Pixi to manage its environment. The Boston Dynamics SDK wheels (v5.1.1) are included in vendor/.

1. Install Pixi by following the installation guide.

2. Clone this repository:

   git clone git@github.com:SaxionMechatronics/spot_rl_inference.git && cd spot_rl_inference

3. Install the dependencies:

   pixi install

4. Convert the training config of the policy you want to deploy. Run the converter and, when prompted, enter the path to the params/ directory of a spot_rl_train checkpoint (the one containing env.yaml). It writes an env_cfg.json next to env.yaml:

   pixi run python spot_rl_runtime/utils/env_convert.py
   # Enter the path to the env.yaml directory
   # /path/to/spot_rl_train/checkpoints/<run>/params

5. Copy the exported policy and the converted config into models/:

   cp /path/to/spot_rl_train/checkpoints/<run>/exported/policy.onnx models/policy.onnx
   cp /path/to/spot_rl_train/checkpoints/<run>/params/env_cfg.json models/env_cfg.json

Usage

Before running, set your robot's credentials and IP address in pixi.toml:

[activation.env]
BOSDYN_CLIENT_USERNAME="user"
BOSDYN_CLIENT_PASSWORD="PutYourPasswordHere"

[tasks]
doyourthing = "python spot_rl_demo.py 192.168.50.3 --policy_file_path ./models -v"

Then start the control loop:

pixi run doyourthing

This connects to Spot, loads the policy from ./models, and begins streaming commands.

Contact

Kousheek Chakraborty - kousheekc@gmail.com

Project Link: https://github.com/SaxionMechatronics/spot_rl_inference

If you encounter any difficulties, feel free to reach out through the Issues section. If you find any bugs or have improvements to suggest, don't hesitate to make a pull request.

Citation

If you use this work in your research, please cite:

@inproceedings{chakraborty2026actuator,
  author    = {Chakraborty, Kousheek and Rajendra, Chandan K. and Alharbat, Ayham and Mersha, Abeje Y.},
  title     = {Actuator Dynamics Curricula for Narrow-Viability Tasks in Legged Robot Learning},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026},
}