embodied-workflows
Guides RLinf embodied training, simulator, real-robot, reward-data, offline-RL, SFT, model, environment, Hydra configuration, and static preflight workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- embodied-workflows
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides RLinf embodied training, simulator, real-robot, reward-data, offline-RL, SFT, model, environment, Hydra configuration, and static preflight workflows.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/embodied-api-and-config.md11.8 KB
- 参考文档references/embodied-recipes.md10.1 KB
- 参考文档references/realworld-and-hardware.md9.2 KB
- 参考文档references/reward-data-and-offline.md12.0 KB
- 参考文档references/troubleshooting.md9.3 KB
- 脚本scripts/check_embodied_config.py25.0 KB
- 脚本scripts/list_embodied_configs.py12.8 KB
03 SKILL.md 正文
RLinf embodied workflows
Use this sub-skill when a task asks about RLinf embodied training setup, simulator or real-world robot recipes, VLA/VLM model and environment YAML choices, reward-model data collection, offline-RL/SFT intersections, or static preflight review before an expensive run.
Do not launch training, evaluation, robot motion, asset downloads, or hardware probes just because this skill is loaded. First identify the target environment family, model family, algorithm, assets, cluster topology, and whether the user wants planning only or an actual run.
Start here
- Classify existing YAMLs with
scripts/list_embodied_configs.pywhen the user gives a config directory or repo root. - Statically inspect a candidate YAML with
scripts/check_embodied_config.pybefore proposing a launch. - Use the recipe and configuration references below to reason about required paths, optional assets, overrides, log paths, video/data-collection settings, and safety constraints.
References
references/embodied-recipes.md— launch pattern, simulator/world-model/real-world recipe families, and model-family pairing guidance.references/embodied-api-and-config.md— embodied runner flow, worker sections, Hydra fields, model/env registries, and static YAML checks.references/realworld-and-hardware.md— Franka/XSquare/DOSW1/GimArm style prerequisites, safety gates, control-node topology, teleoperation, and real-world data collection.references/reward-data-and-offline.md— episode collection, reward-model datasets, VLM Trend rewards, replay buffers, D4RL/offline RL, RECAP/STEAM/CFG, and VLA/VLM SFT intersections.references/troubleshooting.md— simulator assets, EGL/MuJoCo,ROBOT_PLATFORM, optional dependencies, Ray worker failures, data-collection mistakes, reward worker issues, and hardware safety triage.
Routing boundaries
- Ray startup, multi-node environment variables, and placement syntax details belong to
setup-and-cluster; this skill only says which embodied components must be placed and what assumptions to check. - Standalone evaluation, checkpoint resume/inspection, TensorBoard/W&B/SwanLab operations, and debug-log interpretation belong to
operations-evaluation-debugging. - Adding new environment/model source code, registries, workers, tests, Docker, or CI belongs to
extension-development.
Safety defaults
- Prefer static checks and small YAML reviews over real launches.
- Treat real robot movement as a high-risk operation requiring an operator, workspace clearance, e-stop readiness, calibrated cameras/grippers, correct robot IP/serials, and explicit user approval.
- Do not suggest ad-hoc hardware checks on a live robot; use dummy/simulation configs or read-only topology review unless the user explicitly asks for a controlled hardware validation.
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).