How roboparty/ufo works
UFO is a humanoid-control research framework. Its principal workflow is to prepare robot-aware motion data and configuration, train an unsupervised reinforcement-learning policy in the MJLab environment, then run tracking, goal, or reward inference and optionally export an ONNX policy. The graph preserves the README’s data-import and robot-bring-up capabilities as well as the inference branches. Some subsystem internals are unsampled; those links are omitted rather than inferred.
Main entry points: workspace.py Training workspace · tracking_inference.py Tracking inference · reward_inference.py Reward inference · goal_inference.py Goal inference
Components 18
Data and robots
- Robot inspection
humanoidverse/tools/robot_inspect.py - Robot specifications
humanoidverse/utils/robot_spec - Motion data build
humanoidverse/tools/data_build.py - Motion data handling
humanoidverse/utils/motion_data
Training
- Training entrypoint
humanoidverse/train.py - Training workspace
humanoidverse/training/workspace.py - MJLab environment
humanoidverse/agents/envs/humanoidverse_mjlab.py - RL agent algorithms
humanoidverse/agents - Replay buffers
humanoidverse/agents/buffers/trajectory.py - Motion observations
humanoidverse/envs/motion_observations.py
Inference
- Tracking inference
humanoidverse/tracking_inference.py - Goal inference
humanoidverse/goal_inference.py - Reward inference
humanoidverse/reward_inference.py - MJLab inference utilities
humanoidverse/mjlab_inference_utils.py - Checkpoint loading
humanoidverse/agents/load_utils.py - Reward relabeling
humanoidverse/mjlab_reward_relabel.py
Export
- Backward encoder export
humanoidverse/export/backward_encoder.py
Other components
- Engineer
Connections 21
- Engineer → Robot inspection: inspect robot
README.md - Robot inspection → Robot specifications: draft config
humanoidverse/tools/robot_inspect.py - Engineer → Motion data build: build manifest
README.md - Motion data build → Motion data handling: prepare motion data
README.md - Engineer → Training entrypoint: start training
README.md - Training entrypoint → Training workspace: configure training
- Training workspace → MJLab environment: build environment
humanoidverse/training/workspace.py - Training workspace → RL agent algorithms: run agent
humanoidverse/training/workspace.py - Training workspace → Replay buffers: use replay buffer
humanoidverse/training/workspace.py - MJLab environment → Motion observations: compute observations
humanoidverse/agents/envs/humanoidverse_mjlab.py - Engineer → Tracking inference: run tracking
README.md - Engineer → Goal inference: run goal inference
README.md - Engineer → Reward inference: run reward inference
README.md - Tracking inference → MJLab inference utilities: use utilities
humanoidverse/tracking_inference.py - Tracking inference → Checkpoint loading: load checkpoint
humanoidverse/tracking_inference.py - Tracking inference → Backward encoder export: export policy
humanoidverse/tracking_inference.py - Goal inference → MJLab inference utilities: use utilities
humanoidverse/goal_inference.py - Goal inference → Checkpoint loading: load checkpoint
humanoidverse/goal_inference.py - Reward inference → MJLab inference utilities: use utilities
humanoidverse/reward_inference.py - Reward inference → Checkpoint loading: load checkpoint
humanoidverse/reward_inference.py - Reward inference → Reward relabeling: relabel rewards
humanoidverse/reward_inference.py