Nvlabs Compass Compass

You walk a user through onboarding a new robot platform to COMPASS — generating the right config files, registering the embodiment, and verifying the robot spawns correctly. The handbook's extending.md is the canonical written guide; this skill is the operational sibling that does the work interactively.

This skill assumes the docker-as-venv environment from the compass skill — source ./docker/activate already done, container up. If the user hasn't set that up, recommend /compass first.

When NOT to use this skill

• The user wants to switch between built-in embodiments (carter / h1 / spot / g1 / digit) → just use --embodiment <name> on run.py. No code change needed.
• The user is adding a new scene (USD), not a new robot → that's /compass (Register Scene).
• The user has registered a robot but training fails → run /compass-doctor first to surface the actual error.

Workflow

Step 1: Parse what the user already supplied

A user prompt like "add a new robot called Galileo at /workspace/galileo.usd, similar to Spot" already contains three of the four fields the skill needs. Extract them first; only ask about what's missing or genuinely ambiguous. Don't re-ask for things the user already said — that's the dialogue-bot trap.

The four fields:

• Robot name (lowercase identifier like galileo; used in CLI flags and config keys).
• USD path (absolute or repo-relative; the asset that will be spawned).
• Base-class to inherit from (which existing pattern to mirror). See "Base-class choice" below.
• (Optional) display name for docs (e.g., "Galileo").

Step 2: Ask only for what's missing (AskUserQuestion)

If any of the three required fields are missing or ambiguous, ask in one batch. Sample question structure:

For the base-class question, surface the trade-off in the option descriptions:

• wheeled: single base, velocity over base; simplest to onboard. Mirror carter.
• quadruped: articulated legs, ABS-style controller. Mirror spot.
• humanoid: full body; more complex action space. Mirror h1 (or g1/digit for smaller form factors).

Step 3: Read the existing-pattern files

Before generating anything, read the two files that match the user's base-class choice:

Read both top-to-bottom. Don't skim — joint names, action term wiring, and reward shaping all come from this file. Confirming the reference exists before writing keeps you from generating content based on outdated assumptions.

Step 4: Show the diff before writing

Generate a preview of what will be created or modified:

1. New ArticulationCfg block to be appended to compass/rl_env/exts/mobility_es/mobility_es/config/robots.py. Mirror the matching reference, swap robot name + USD path + joint config.
2. New file: compass/rl_env/exts/mobility_es/mobility_es/config/<robot_name>_env_cfg.py. Mirror the matching reference; swap class name (e.g., <RobotName>GoalReachingEnvCfg), scene.robot to point at the new ArticulationCfg, action term to whatever joint controller fits.
3. Edit run.py:
   • Add import: from mobility_es.config.<robot_name>_env_cfg import <RobotName>GoalReachingEnvCfg.
   • Add row to EmbodimentEnvCfgMap: '<robot_name>': <RobotName>GoalReachingEnvCfg, (around line 116–122).
4. Optional gin update: if the user wants <robot_name> to be the default embodiment, edit configs/shared.gin's embodiment field. By default, leave existing default in place and let the user pass --embodiment <robot_name> on the CLI.

Show the user each preview block (or unified diff). Ask for confirmation before writing.

Anti-pattern guard: never silently write multi-file scaffolding. Robot onboarding is high-blast-radius — wrong joint counts or an inverted action sign can cause the robot to flail into the floor in simulation. The diff-then-confirm gate is non-negotiable.

Step 5: Apply the changes

After confirmation, write the files and edit run.py. Use the Edit tool for run.py (one Edit per insertion: import + map row). Use Write for the new env_cfg.py. For robots.py, use Edit to append the new ArticulationCfg block at the end (read the file first to find the right insertion point — the last existing config).

Step 6: Smoke test

Verify the robot spawns by running training with --num_envs 1 and without --headless so the user can see the robot appear:

python run.py \
  -c configs/train_config.gin \
  --enable_cameras \
  -o /tmp/<robot_name>_smoke \
  -b ./assets/x_mobility.ckpt \
  --logger tensorboard \
  --embodiment <robot_name> \
  --environment warehouse_single_rack \
  --num_envs 1

Use run_in_background: true. Tell the user to watch the GUI:

1. Robot spawns at the configured initial pose (typically (0, 0, ~1.0)).
2. Robot doesn't immediately fall through the floor (z-init too low) or shoot up (z-init too high).
3. First PPO iteration completes without joint-controller errors in the log.

If the robot misbehaves, route to /compass-doctor for a deeper diagnostic. Common first-onboarding issues:

• Z-init too low → robot clips through floor → bump init_state.pos[2].
• Joint names mismatch → ArticulationCfg fails to load → recheck against the USD's joint hierarchy.
• Action term sign flipped → robot drives backwards → invert sign in the action term.

Step 7: Document the new embodiment (optional)

Once smoke test passes, suggest the user add their robot to docs/handbook/extending.md (under "Built-in embodiments" or as a new subsection). This is optional — many users keep their custom embodiment private. Don't push the doc change if the user is in a hurry; just flag it for later.

Reference: file map

Reference: base-class examples

• Wheeled (carter): see carter block at robots.py:23 and carter_env_cfg.py.
• Quadruped (spot): see spot block at robots.py:275 and spot_env_cfg.py.
• Humanoid (h1): see h1 block at robots.py:72 and h1_env_cfg.py.

For canonical written guidance (joint config, controller selection, reward shaping), the handbook's extending.md covers depth this skill won't replicate.

Source: NVlabs/COMPASS — distributed by TomeVault.