training-and-evaluation
Train, evaluate, and export HumanoidVerse policies with PPO, PPODeltaA, Hydra overrides, simulator selection, motion tracking, locomotion, and delta-action workflows.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training-and-evaluation
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Train, evaluate, and export HumanoidVerse policies with PPO, PPODeltaA, Hydra overrides, simulator selection, motion tracking, locomotion, and delta-action workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md14.7 KB
- 参考文档references/troubleshooting.md7.2 KB
- 参考文档references/workflows.md16.0 KB
- 脚本scripts/build_training_command.py17.4 KB
03 SKILL.md 正文
training-and-evaluation
Use this sub-skill when a user asks for HumanoidVerse policy training, checkpoint evaluation, ONNX export, simulator/backend choice, Hydra override selection, phase-based motion tracking, locomotion, open-loop delta-action training, or closed-loop delta-action finetuning.
Run ASAP training/evaluation entry points from the repository checkout root. The bundled helper scripts are skill-root-relative and accept --repo-root <asap-checkout>. The repo entry points are scripts, not package modules: python humanoidverse/train_agent.py ... and python humanoidverse/eval_agent.py .... Do not use python -m humanoidverse.train_agent, because the source imports utils.config_utils relative to the humanoidverse/ script directory.
Do not use this sub-skill for
- SMPL/AMASS shape fitting, motion retargeting, robot XML fitting, or MuJoCo retargeting visualization. Use
../motion-retargeting/SKILL.md. - Sim2sim or sim2real runtime control, ROS2/Unitree bridges, joystick control, and live robot deployment. Use
../sim2real-deployment/SKILL.md. - Generic PPO theory or unrelated simulator stacks that are not launched through this repository's HumanoidVerse Hydra configs.
Required reading order
- Read the root router:
../../SKILL.md. - Read root install/backend guidance before choosing a simulator:
../../references/install-and-backends.md. - Read this sub-skill's command recipes:
references/workflows.md. - Read CLI/config/class behavior before changing Hydra groups or checkpoint/export assumptions:
references/api-reference.md. - If anything fails, read this sub-skill's predictable failures first:
references/troubleshooting.md, then root troubleshooting../../references/troubleshooting.md.
Safe preflight checks
These checks do not start simulator training:
# Confirm that Hydra can list config groups.
python humanoidverse/train_agent.py --help
python humanoidverse/eval_agent.py --help
# Build a command without executing it.
python sub-skills/training-and-evaluation/scripts/build_training_command.py \
--repo-root <asap-checkout> \
--workflow motion-tracking \
--motion-file humanoidverse/data/motions/g1_29dof_anneal_23dof/TairanTestbed/singles/0-TairanTestbed_TairanTestbed_CR7_video_CR7_level1_filter_amass.pkl
# Turn a training command into a Hydra composition-only smoke check.
python sub-skills/training-and-evaluation/scripts/build_training_command.py \
--repo-root <asap-checkout> \
--workflow motion-tracking \
--motion-file humanoidverse/data/motions/g1_29dof_anneal_23dof/TairanTestbed/singles/0-TairanTestbed_TairanTestbed_CR7_video_CR7_level1_filter_amass.pkl \
--cfg-job
Run the root doctor script before runtime commands to get a single import/help report:
python scripts/asap_doctor.py --repo-root <asap-checkout> --section core
Route by prompt
| User asks for | Use | Primary recipe |
|---|---|---|
| "train motion tracking", "imitate this motion", "CR7 policy" | train_agent.py with +exp=motion_tracking |
references/workflows.md#motion-tracking-training |
| "train locomotion" or "test IsaacGym install with locomotion" | train_agent.py with +exp=locomotion |
references/workflows.md#locomotion-training-and-smoke-test |
| "train delta action model" | train_agent.py with +exp=train_delta_a_open_loop |
references/workflows.md#open-loop-delta-action-training |
| "finetune with delta A" or "closed-loop delta action" | train_agent.py with +exp=train_delta_a_closed_loop and algo.config.policy_checkpoint=... |
references/workflows.md#closed-loop-delta-action-finetuning |
| "evaluate checkpoint", "visualize policy" | eval_agent.py +checkpoint=... |
references/workflows.md#checkpoint-evaluation |
| "export ONNX" | eval_agent.py +checkpoint=...; export happens before the eval loop |
references/workflows.md#onnx-export |
| "which Hydra overrides?" | Choose config groups and scalar overrides from the workflow table | references/api-reference.md#hydra-config-groups |
Backend selection rules
- Default to
+simulator=isaacgymfor README-aligned training recipes when IsaacGym Preview 4 is installed and CUDA is available. - Use
+simulator=isaacsimonly in an IsaacLab/IsaacSim environment withomni.isaac.lab; the entry points launch an IsaacSim app before importing Torch. - Use
+simulator=genesisonly after installinggenesis-worldin the same environment. +simulator=mujocois present as a Hydra config, but deployment/control recipes live in../sim2real-deployment/SKILL.md; do not infer that every training recipe has been validated on MuJoCo.- The verified inspection environment had CUDA-capable Torch, Hydra,
humanoidverse,asapdistribution metadata,onnx, andonnxruntime, but did not have IsaacGym, Genesis, or IsaacLab installed. Treat actual simulator launch as backend-gated.
Common outputs
- Training experiment directory:
logs/<project_name>/<timestamp>-<experiment_name>-<log_task_name>-<robot_type>/. - Training checkpoints:
model_<iteration>.ptin the experiment directory;PPO.save()writes actor, critic, optimizer states, iteration, and infos. - Saved training config:
config.yamlin the experiment directory; Hydra's own run directory is<experiment_dir>/.hydra/andtrain.logis written there. - Training renderings:
<experiment_dir>/renderings_training/when the simulator records them. - Evaluation logs:
logs_eval/<eval_name>/<eval_timestamp>/config.yamlandeval.logunder the Hydra eval run directory. - Evaluation renderings:
<checkpoint_dir>/renderings/ckpt_<N>/, where<N>comes frommodel_<N>.pt. - ONNX export:
<checkpoint_dir>/exported/model_<N>.onnx;eval_agent.pyexports before it enters the infinite evaluation loop. - Evaluation motion recording with
+opt=record:<checkpoint_dir>/motions/<save_note>_<eval_timestamp>.pklor<eval_timestamp>_<dump_motion_name>.pkl; those dumps include anactionkey useful for delta-action datasets.
Cross-links
- Root router:
../../SKILL.md. - Root install and backend notes:
../../references/install-and-backends.md. - Root troubleshooting:
../../references/troubleshooting.md. - Motion retargeting:
../motion-retargeting/SKILL.mdfor producing motion.pklfiles consumed by motion-tracking training. - Sim2real deployment:
../sim2real-deployment/SKILL.mdfor using exported ONNX policies in MuJoCo/ROS2/Unitree runtime flows.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).