[TomCC7]

MakerMods Metal stations can collect demonstrations and run Dreamlabs EEF policies through the shared learning controls, using one robot model for teleop, planning, policy FK, and execution. This integration includes the Metal implementation and the Dreamlabs backend from #4653; no additional merge of #4865/#4881 is required.

Collection and rollout blueprints cover single Metal with overhead, dual Metal with two wrists, and dual Metal with overhead plus wrists. Live testing confirmed the public dreamlabs/dream-1-beta checkpoint requires all three camera roles; the smaller layouts remain available for collection and compatible custom checkpoints. Metal accepts EEF output only, and Dreamlabs joint output remains restricted to canonical dual YAM. Preview is the default, dispatch is explicit, and Quest grip takeover stops policy without automatic resume.

The shared model now distinguishes the execution grasp_frame (80 mm along gripper_base +Z) from the Dreamlabs protocol dreamlabs_tcp (135 mm along vendor link6 +X, RPY pi/2, 0, pi/2). The protocol frame matches the actual deployed MakerMods playground policy worker. Using grasp_frame directly previously introduced a fixed 20.8 mm and 90-degree discrepancy. Observations use dreamlabs_tcp in each arm's own base frame, and returned protocol poses are transformed back to grasp_frame. Both links are defined in Python, with side prefixes for dual Metal; no separate TCP JSON is required for the standard model. Named or explicit legacy tool-transform overrides remain supported.

Physical Metal arms enable the shared adapter's URDF-based gravity compensation. This stack also includes EEF CBOR RPC serialization, coordinator task-call compatibility fixes, a bounded wait for transient observation skew, the SDK's 15-second cloud request timeout, and detailed trajectory fault reasons. Metal uses a 5 cm final position tolerance. Optional station debugging settings can disable endpoint convergence or observation timing gates; their defaults remain enabled, and hardware limits, finite-value checks, tracking bounds, and deadman supervision remain active.

Validation of the protocol-frame correction: 72 focused regressions and 8 cached-URDF/Pink/mock tests passed, plus scoped mypy and pre-commit. Golden FK vectors were generated offline from the actual deployed playground policy-worker-DxdwLh8C.js, with its URL and SHA256 recorded in the fixture. Tests cover outgoing state and incoming action conversion, single/dual arms, nontrivial base mounts, and existing override compatibility. The deployed worker's FK and the corrected model agree within 1e-7 m in these checks. No new inference or motor commands were issued to validate this correction.

Earlier station testing reached live three-camera inference and hardware EEF execution, and exposed tracking and convergence faults. Sustained task success, physical TCP calibration, measured dual mounting geometry, and real execution after the protocol-frame correction remain unverified. Default dual mounts are a mock example, not measured installation geometry.

This remains an integration draft based on #4653. Station guide: docs/capabilities/manipulation/learning/guides/metal-dreamlabs-setups.md.

[github-actions]

go2 replay realtime (arm64)

Details

Benchmark suite	Current: 8d960d6	Previous: 6c5624a	Ratio
peak memory	1752.547 MB	1779.605 MB	0.98
peak threads	365 threads	374 threads	0.98
network (transport)	2726.236 MB	2736.27 MB	1.00
disk write	14.324 MB	1.496 MB	9.57
instructions	137.723 G	138.686 G	0.99

This comment was automatically generated by workflow using github-action-benchmark.

[codecov]

❌ 1 Tests Failed:

Tests completed	Failed	Passed	Skipped
7653	1	7652	57

View the top 1 failed test(s) by shortest run time

dimos.codebase_checks.test_no_underscore_assign::test_no_underscore_assignment

Stack Traces | 8.09s run time

def test_no_underscore_assignment():
        """Fail if any file assigns to a bare `_`."""
        dimos_dir = DIMOS_PROJECT_ROOT / "dimos"
        hits = find_underscore_assignments()
        if hits:
            listing = "\n".join(f"  - {p.relative_to(dimos_dir)}:{lineno}" for p, lineno in hits)
>           raise AssertionError(
                f"Found assignment(s) to `_` in dimos/:\n{listing}\n\n"
                "Assigning to `_` is not allowed: it hides an unused variable instead "
                "of removing it. Delete the variable. If you only need the "
                "expression's side effect, evaluate it directly with a call "
                "(`obj.method()`, `getattr(obj, 'attr')`) or log it; a bare attribute "
                "access needs `# noqa: B018`. Tuple unpacking (`a, _ = f()`) is fine "
                "and not flagged by this rule."
            )
E           AssertionError: Found assignment(s) to `_` in dimos/:
E             - control/test_task_context.py:70
E             - control/test_task_context.py:113
E             - control/test_task_context.py:128
E           
E           Assigning to `_` is not allowed: it hides an unused variable instead of removing it. Delete the variable. If you only need the expression's side effect, evaluate it directly with a call (`obj.method()`, `getattr(obj, 'attr')`) or log it; a bare attribute access needs `# noqa: B018`. Tuple unpacking (`a, _ = f()`) is fine and not flagged by this rule.

dimos_dir  = PosixPath('.../dimos/dimos/dimos')
hits       = [(PosixPath('.../dimos/dimos/dimos/control/test_task_context.py'), 70), (PosixPath('.../runner/work/...ol/test_task_context.py'), 113), (PosixPath('.../dimos/dimos/dimos/control/test_task_context.py'), 128)]
listing    = '  - control/test_task_context.py:70\n  - control/test_task_context.py:113\n  - control/test_task_context.py:128'

dimos/codebase_checks/test_no_underscore_assign.py:55: AssertionError

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}