perception-terrain-randomization
Configure mjlab sensors, terrain generation, flat patches, and domain randomization with CPU/CUDA verification boundaries.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- perception-terrain-randomization
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Configure mjlab sensors, terrain generation, flat patches, and domain randomization with CPU/CUDA verification boundaries.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/domain-randomization.md10.4 KB
- 参考文档references/sensors.md10.8 KB
- 参考文档references/terrain.md9.7 KB
- 参考文档references/troubleshooting.md2.5 KB
- 脚本scripts/inspect_sensor_terrain.py2.6 KB
03 SKILL.md 正文
Perception, Terrain, and Randomization
Use this sub-skill when an mjlab task needs perception sensors, terrain, terrain-aware resets/curriculum, or runtime domain randomization.
Route here for
- Scene-level
BuiltinSensorCfg,ContactSensorCfg,RayCastSensorCfg,TerrainHeightSensorCfg, andCameraSensorCfgselection. - Sensor data semantics: contact reductions/history, raycast hit distances, terrain height scans, RGB/depth/segmentation outputs, and debug visualization.
TerrainEntityCfg,TerrainGeneratorCfg, terrain presets, curriculum grid behavior, spawn proportions, and flat patch sampling.- Domain-randomization event functions in
mjlab.envs.mdp.dr, including model field expansion, recomputation levels, camera/render implications, and visualization limits. - Deciding whether a check is CPU-config-only or a CUDA/full sensor validation.
Route elsewhere
- Base entity, scene, MuJoCo spec editing, asset variants, and scene export: scene-simulation-assets.
- Observation/reward/action/termination term dictionaries and observation-group wiring: environment-configuration and mdp-components.
- Training, playback, task registry, viewer CLI, checkpoints, W&B, and export commands: training-evaluation-cli.
Operating workflow
- Identify which layer owns the request: sensor config, terrain config,
terrain-aware reset/curriculum, or
drevent term. - Confirm exact installed signatures and terrain preset names with the bundled
inspection script when the mjlab version
may matter. Use
--jsonfor machine-readable output. - Read the relevant reference:
- Sensors for built-in/contact/raycast/camera data.
- Terrain for generator modes, presets, flat patches, curriculum, and debug groups.
- Domain randomization for event modes, field selection, operations/distributions, and recompute behavior.
- Troubleshooting for common sensor, terrain, rendering, and DR failures.
- Choose verification level explicitly:
- CPU/config checks can inspect imports, dataclass signatures, CLI/help text, and pure terrain config construction.
- CUDA/full checks are needed for raycast/camera sensing, MuJoCo Warp render context behavior, sensor-context creation, and training-scale simulation.
Fast decisions
- Use
ContactSensorCfgfor structured contact data and gait air-time signals. - Use
RayCastSensorCfgwith aGridPatternCfgfor terrain scans and with aPinholeCameraPatternCfgfor depth-like rays. - Use
TerrainHeightSensorCfgwhen the policy needs per-frame vertical clearance rather than raw raycast hit positions. - Use
CameraSensorCfgfor RGB/depth/segmentation images; keep all camera render settings identical inside one scene. - Use terrain named configs for standard locomotion terrain, then customize with
dataclasses.replace()or by replacingsub_terrains. - Prefer
startuporresetevents for DR functions that trigger model recomputation; reserve frequent interval/step events for cheap state or friction-style changes.
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).