AMP_unilab

中文说明

AMP G1 locomotion/recovery training migrated from AMP_mjlab (source revision 6c7a2947) onto the UniLab Manager-Based runtime. UniLab and unilab-rl are external pinned dependencies; this package owns only task configuration, AMP public-runtime bridging, and the vendored AMP-capable RSL-RL runner.

Tasks

• Unitree-G1-AMP-Flat (MuJoCo Warp backend by default; CPU mjbatch remains available with --sim mujoco)
• 29 controlled joints
• actor observation: 96 terms × 4 history = 384
• critic observation: 216 terms × 4 history = 864
• AMP discriminator state: 13 bodies × 15 = 195; transitions use 390 inputs
• 20 s episode, 50 Hz physics, 25 Hz control, expert clips: all 18 walk/run and recovery NPZ files

UniLab discovers the task through the unilab.tasks entry point:

amp-list-envs
amp-train --task Unitree-G1-AMP-Flat --num-envs 4096
amp-export --checkpoint runs/g1_amp_unilab/model_final.pt --output policy.onnx
amp-play --checkpoint runs/g1_amp_unilab/model_final.pt --steps 1000

The source G1 AMP PPO/AMP hyperparameters are preserved in src/amp_unilab/training/config.py: 24 steps/env, four mini-batches, adaptive learning rate, 0.75 task-reward lerp, 0.1 AMP reward coefficient, 100k AMP replay buffer, [1024,512,256] discriminator, and [512,256,128] actor/critic MLPs with empirical observation normalization.

Robot assets

The G1 structural XML is repository-owned at src/amp_unilab/assets/robots/g1/g1.xml; the flat scene and task keyframe are in the adjacent scene_flat.xml. Mesh STLs stay out of git. At configuration time, the package writes a content-addressed runtime XML whose meshdir points to UniLab's registered asset hub.

Commands

uv sync
make check        # formatting and lint
make test         # fast registry/config/motion tests
make test-all     # also constructs MuJoCo and runs a one-iteration CPU AMP update
make build

Migration status and known differences

The objective is behavioral parity, but this first migration is deliberately explicit about what differs instead of hiding it behind compatibility seams.

Aligned or closely bridged

• Flat G1 task, command model, control rate, action ordering and exact scales
• actor/critic term composition and dimensionality
• AMP state construction and all 18 expert clips
• expert transition sampling and AMP discriminator/optimizer behavior
• motion reset sampling from walk/run groups, with recovery clips selected for the source's 40% delayed-reset subset
• the repository-owned G1 structure, task keyframe, Unitree PD gains, armature, and torque-speed curve are control-step equivalents of AMP_mjlab's UnitreeActuator semantics
• play mode disables observation noise and interval pushes and samples recovery resets for all environments

Backend difference

The default MuJoCo Warp path omits push_robot because UniLab's interval root-velocity-delta capability is unavailable on that backend. The task keeps this difference explicit instead of probing backend internals; use --sim mujoco when interval velocity pushes are required.

Current differences

• History remains UniLab's term-major representation. AMP_mjlab uses time-major [T-3,...,T]; UniLab's public observation manager currently concatenates history term-major.
• AMP_mjlab evaluates its actuator model on every physics substep. UniLab's public per-substep action feedback currently cannot refresh tracked body state with MuJoCo implicitfast, so AMP_unilab preserves the source integrator and evaluates the torque-speed/friction model once per control step.
• The source DelayedTerminationManager (40% of environments, up to 250 control steps) is not installed because replacing a manager is outside UniLab's declared manager lifecycle. Consequently the delayed-root-height masking and 3.5× delayed reward path cannot be reproduced yet.
• Startup domain randomization for foot friction, encoder bias, and base COM offset is not enabled; the corresponding UniLab MuJoCo capability bridge is not part of this task package.
• Foot slip and self-collision penalties require contact force/history capabilities and are not present.
• The three-stage velocity command curriculum and rough terrain/height scan are not present. Only Flat is exposed.

Training results before closing these gaps are not expected to be numerically identical to AMP_mjlab. Do not treat this as a benchmark-compatible release.

License

Task files follow the repository's BSD-3-Clause licensing. Vendored src/amp_unilab/vendor/rsl_rl retains its original upstream notices and BSD-3-Clause license.