ttrl-isaaclab-editor

name	ttrl-isaaclab-editor
description	Use when modifying, extending, debugging, or reviewing the TTRL IsaacLab table-tennis repository at /shared_disk/users/fanyahao/TTRL, especially for task registration, TTEnv or G1Planner env behavior, rewards, observations, actions, resets, serving logic, physics, assets, play or eval variants, latent-skill integration, RSL-RL runners, predictor integration, or checkpoint-sensitive task variants.

TTRL IsaacLab Editor

Route edits in TTRL to the smallest correct ownership layer before changing code.

Required Load Order

• Treat /shared_disk/users/fanyahao/TTRL as the source of truth.
• Read repo instructions and local task notes before editing.
• Start with selection-guide.md.
• Then load edit-layer-decision-tree.md.
• Then load exactly one recipe that matches the request.
• Then load exactly one source note for the first-owner layer.
• Only load external-isaaclab.md if the local repo and this skill are insufficient.

Load Recipes By Intent

• Existing reward, observation, action, reset, serve, or eval behavior edit -> task-customization.md
• New task variant, eval variant, or config variant -> task-authoring.md
• Runner, policy, predictor, latent-skill module, checkpoint, resume, export, or training bridge change -> training-extension.md
• Validation, compatibility, or completion checks -> validation-checklists.md

Load Source Notes By Ownership

• Task ids, task registry, train/eval config resolution -> task-registration.md
• Shared table-tennis env base, reset loop, observation assembly, serve lifecycle, prediction bridge -> env-base.md
• Planner task family, planner-specific cfg, right-arm latent-skill contract, eval variant -> g1-planner-task.md
• Reward, event, curriculum, and serve-randomization helpers -> mdp.md
• Ball aerodynamics and physics-side force integration -> physics.md
• Robot, ball, and table asset contracts -> assets.md
• RSL-RL runners, latent-skill policy modules, and training-side action contract -> rsl-rl.md
• CLI flags, wrappers, play/eval/train entrypoints, and predictor routing -> scripts.md

Working Rules

• Anchor to g1_planner and g1_planner_eval, not the older t1_tt examples still present in TTRL/README.md.
• Prefer the narrowest owner: task cfg or task env before shared base; local wrapper before runner; local runner before generic module.
• If an observation or action contract changes, inspect TTEnv or G1PlannerEnv, LeggedLabGymEnvAdapter, LeggedLabRslRlVecEnvWrapper, agent cfg, and runner compatibility in the same work item.
• If checkpoint compatibility may break, prefer a new task id or explicit variant rather than silently mutating the existing training path.
• Use symbol search when paths drift: task_registry.register, G1PlannerEnvCfg, ball_reset_ids, update_prediction, skill_resume_path, OnPolicyPredictorRegressionRunner.