feat(sim): add xarm6-worldbelief-sim blueprint

[mustafab0]

Motivation

The WorldBelief perception stack can only be exercised with an xArm6 on the desk. This makes the sim a first-class dev surface: dimos run xarm6-worldbelief-sim brings up the same perception/planning/MCP stack against MuJoCo, so scan, recall, TF, gripper, and the agent path can be developed and regression-tested without hardware.

What it wires

xarm6-worldbelief-sim mirrors xarm6-worldbelief with the RealSense wrist camera replaced by MujocoSimModule (wrist_camera, base_frame_id=link6, dof=6, XARM6_SIM_PATH). It renders the MJCF camera and publishes the same color/depth/camera-info topics plus link6 -> wrist_camera_*_optical_frame from sim ground truth; world -> link6 comes from ManipulationModule FK.

• The rerun bridge, recorder, belief module, and MCP server were byte-identical between the two blueprints, so they move into a worldbelief_stack(state_subdir) factory. xarm6_worldbelief is otherwise untouched — verified by diffing every atom's module + kwargs before/after (only function object addresses differ).
• XARM6_SIM_HOME puts wrist_camera at world (0.45, 0, 0.52) looking straight down. Solved against data/xarm6/scene.xml as shipped, i.e. with link_base on the 0.12 m pedestal — the same convention XARM7_SIM_HOME already follows (verified: main's xarm7 home lands the camera at 0.44991, 0, 0.51979).
• make_xarm6_sim_hardware / make_xarm6_sim_module_kwargs mirror the existing xarm7 factories.

No changes outside blueprint, config, registration, and tests.

Coordination with #3275 (worldbelief → experimental)

#3275 (approved) relegates worldbelief to dimos/experimental/world_belief/, including moving xarm/blueprints/worldbelief.py → experimental/world_belief/xarm6_blueprint.py.

By design this PR stays green against main and does not pre-empt those paths. worldbelief_sim.py intentionally stays in the xarm package. Once #3275 lands, this blueprint needs exactly one import line changed:

-from dimos.robot.manipulators.xarm.blueprints.worldbelief import worldbelief_stack
+from dimos.experimental.world_belief.xarm6_blueprint import worldbelief_stack

plus the same substitution for the two lazy imports in test_worldbelief_sim_blueprint.py. Note #3275 renames worldbelief.py at 96% similarity, so its move and this PR's edits to that file will need a merge resolution whichever order they land in.

Rebase notes (onto #3183)

Rebased onto current main. Two API changes absorbed:

• dimos/robot/manipulators/test_blueprints.py was deleted by feat: extract manipulation plan execution manager #3183, so the blueprint-composition tests moved to a scoped dimos/robot/manipulators/xarm/test_worldbelief_sim_blueprint.py rather than resurrecting the deleted file.
• RobotModelConfig.coordinator_task_name was removed by feat: extract manipulation plan execution manager #3183; the test now derives the expected task name from trajectory_task_name(robot.name).
• xarm6-planner-only was dropped from SELF_HOSTED_BLUEPRINTS on main; that removal is preserved.

Validation

Re-run after the rebase, so these reflect #3183's new plan-execution manager.

	result
Launch	7/7 modules deploy, 0 errors; camera transports wired; arm reaches home
Trajectory	move_to_joints → "Reached target joint configuration"; 51-point trajectory completes; arm converges toward target (0.530 → 0.546 → 0.558 → 0.6)
Scan	"Scan complete: 1 object(s) present" (49 s, 12 frames folded)
MCP	server lists 14 skills; go_home / move_to_joints / get_robot_state / scan all invoke
Agent	McpClient deploys and discovers all 14 tools over MCP
Tests	280 pass (blueprint suites + generation + kwargs checks); lint/format clean

Detected red ball vs MJCF ground truth (apple at 0.40, 0.08, 0.17):

axis	detected	truth	error
X	0.4094	0.40	+0.94 cm
Y	0.0697	0.08	−1.03 cm
Z	0.0746	0.17	−9.54 cm

Known issues (the PR ladder)

Documented, not fixed — each is its own rung.

1. Base-frame disagreement → scan Z reads ~9.5 cm low; pick will fail IK. data/xarm6/scene.xml spawns link_base at pos="0 0 .12" (plus a pedestal geom), but make_xarm_model_config places base_pose at the world origin. The sim renders from the true camera pose while perception unprojects using the FK pose. Measured −9.54 cm above, matching prediction. Scan/detect/recall work; pick does not. Next rung.
2. Gripper direction inverted. hand.xml's driver joint grows as the jaws close; DiMOS positions grow as they open. Needs gripper_position_inverted support in MujocoSimModule, which main does not have.
3. Pink IK aborts retries on QP exception. Needed before pick converges.
4. scan / recall exceed the 120 s default RPC timeout. Server-side the skill succeeds (SKILL scan result=OK duration_ms=156804.7) but the MCP layer reports an error at 120 s. Aggravated by (5).
5. Detector runs on CPU on this box — Failed to initialize NVML: Driver/library version mismatch (NVML 595.84), so torch.cuda.is_available() is False. Environment, not code, but it is what pushes scan past the RPC limit.
6. dimos mcp call hardcodes a 30 s HTTP read timeout, below the scan's runtime even on the happy path.
7. Trajectories converge asymptotically. move_to_joints/go_home report success while the arm is still travelling; sampling immediately after the call shows it near its start. Repeated go_home calls converge (joint 0: 0.417 → 0.021 → 0.013 → 0.008).
8. Scan prompt sensitivity. With ["red ball", "orange ball", "cup"] only the red ball is detected; the orange and cup are missed.

Not verified

The agent's natural-language exchange could not complete: the available OpenAI key is rejected with 401 invalid_api_key (confirmed with a direct curl to api.openai.com, so it is the credential, not the plumbing). Everything up to the LLM call works — the agent deploys, connects, and discovers all 14 MCP tools.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

@@            Coverage Diff             @@
##             main    #3242      +/-   ##
==========================================
+ Coverage   75.20%   75.22%   +0.01%     
==========================================
  Files        1128     1130       +2     
  Lines      107954   108020      +66     
  Branches     9751     9752       +1     
==========================================
+ Hits        81188    81254      +66     
- Misses      23955    23956       +1     
+ Partials     2811     2810       -1     

Flag	Coverage Δ
OS-ubuntu-24.04-arm	68.85% <56.89%> (-0.01%)	⬇️
OS-ubuntu-latest	70.91% <56.89%> (-0.01%)	⬇️
Py-3.10	70.91% <56.89%> (-0.02%)	⬇️
Py-3.11	70.90% <56.89%> (-0.02%)	⬇️
Py-3.12	70.91% <56.89%> (-0.01%)	⬇️
Py-3.13	70.91% <56.89%> (-0.01%)	⬇️
Py-3.14	70.91% <56.89%> (-0.01%)	⬇️
Py-3.14t	70.90% <56.89%> (-0.02%)	⬇️
SelfHosted-Large	29.10% <29.62%> (+0.01%)	⬆️
SelfHosted-Linux	35.64% <83.58%> (-0.02%)	⬇️
SelfHosted-macOS	34.71% <83.58%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
dimos/robot/all_blueprints.py	100.00% <ø> (ø)
.../robot/manipulators/xarm/blueprints/worldbelief.py	87.09% <100.00%> (+1.38%)	⬆️
...ot/manipulators/xarm/blueprints/worldbelief_sim.py	100.00% <100.00%> (ø)
dimos/robot/manipulators/xarm/config.py	90.47% <100.00%> (+0.82%)	⬆️
...anipulators/xarm/test_worldbelief_sim_blueprint.py	100.00% <100.00%> (ø)
dimos/robot/test_all_blueprints.py	87.50% <ø> (ø)

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.