Sumo 🥋

Disclaimer: this code is released accompanying a paper submission only. The paper itself can be found here.

sumo🥋 is a research codebase for whole-body loco-manipulation built on judo🥋.

Install

Requires pixi.

# Install the default app environment
pixi install

# Build the local g1_extensions extension (required for G1 tasks)
# and judo's mujoco_extensions (required for Spot tasks)
pixi run build

pixi run sumo assumes both native extensions above have already been built.

If you prefer named environments, pixi install -e dev is equivalent for this repo, and the corresponding task form is pixi run -e dev ....

Run

# Launch simulation + visualizer with default task
pixi run sumo

# Specify a task
pixi run sumo task=spot_box_push

# Specify a task and other hydra options (e.g. optimizer)
pixi run sumo task=g1_box optimizer=mppi

Headless MPC Runner

run_mpc runs MPC episodes without the full GUI, useful for benchmarking and data collection.

# Run with defaults (spot_box_push, CEM optimizer, 2 episodes)
pixi run python -m sumo.run_mpc

# Specify task and options
pixi run python -m sumo.run_mpc --init-task=g1_door --init-optimizer=cem --num-episodes=10

# With visualization
pixi run python -m sumo.run_mpc --init-task=g1_box --visualize

# Disable saving results
pixi run python -m sumo.run_mpc --init-task=spot_box_push --no-save-results

# Record all rollout data
pixi run python -m sumo.run_mpc --init-task=g1_door --record-all-data

Results are saved to run_mpc/results/ by default as HDF5 files. Use --help to see all options.

Tests

pixi run pytest tests/ -v

Reference

If you find this codebase helpful, please consider citing the SUMO paper.

@article{zhang2026sumo,
  title={Sumo: Dynamic and Generalizable Whole-Body Loco-Manipulation},
  author={John Z. Zhang and Maks Sorokin and Jan Brudigam and Brandon Hung and Stephen Phillips and Dmitry Yershov and Farzad Niroui and Tong Zhao and Leonor Fermoselle and Xinghao Zhu and Chao Cao and Duy Ta and Tao Pang and Jiuguang Wang and Preston Culbertson and Zachary Manchester and Simon Le Cleac'h},
  year={2026},
  note={TODO: add arXiv link}
}