Factory: nine tasks, one matched expert

Source audit and evaluation design. This page contains no Factory experiment outcomes. All nine tasks use the native20-second horizon. Native success remains authoritative, with the proxy limitations below disclosed.

The public benchmark supplies100 simulation demonstrations per task. The published GR00T-N1.7 checkpoint covers all nine; ACT receives the corresponding known task ID. Expert card · Dataset card · Full reset ranges and provenance · Frozen evaluation protocol

Task and instruction	Native success predicate	Evidence
LMBoxLift Walk to the table and pick up the box	Object root world Z >0.18m and at least one finger link within0.2m.	Pinned source
LMBoxLiftFloor Walk to the box and pick it up from the floor	Object root world Z >−0.5m, local-Z/world-up dot ≥0.9, and at least one finger within0.6m.	Pinned source
LMBoxTableToShelfPnP Move the box from the table to the shelf	Object/destination root separation per axis <[0.16,0.14,0.9227]m, object static, every finger >0.1m away, and destination contact >0.95N.	Pinned source
LMDrillLift Walk to the table and pick up the power drill	Drill root world Z >0.16m and approximate grasp predicate.	Pinned source
LMDrillLiftObs Navigate around the shelf and pick up the power drill from the table	Drill root world Z >0.16m and approximate grasp predicate. Success does not explicitly require obstacle avoidance.	Pinned source
LMDrillPnP Move the power drill from one table to another	Drill/destination-table root separation <[0.3,0.27,0.81]m, static, local-Z/world-up dot ≥0.8, and every finger >0.1m away. No contact term.	Pinned source
LMPickDrillFromHolder Pick up the power drill from the holder	Drill root world Z >0.13m, approximate grasp, and holder contact is not above0.95N.	Pinned source
LMPushButton Walk to the console and press the red button	Button joint normalized pressedness >0.5 and at least one finger within0.2m.	Pinned source
LMPushShelfForward Push the tall wheeled shelf to the marked area	Shelf/target root separation <[0.3,0.3,0.01]m, static, local-Z/world-up dot ≥0.9, and a finger within1.1m.	Pinned source

Interpretation and implementation risks

Native lift criteria use absolute object-root world Z, not displacement from reset. Root origins can be below the visible object.
Approximate grasp means any of14 finger links within0.18m plus at least3 finger joints displaced >0.1rad from defaults; no required object-hand contact sensor, and proximity/closure are not explicitly bound to the same hand.
Static means linear speed ≤0.15m/s and angular speed ≤0.15rad/s. Placement uses per-axis world root separation, not a support-surface polygon.
A proximity predicate or a single terminal frame is not proof of a secure sustained grasp. The collection CLI success-hold option does not automatically add dwell to inference success.
The obstacle task instruction requests avoidance, but its success predicate only tests drill lift/grasp. Keep collisions and videos visible rather than reinterpret native success.
Button reset_pressedness=0.8 maps through the Arena unpress event to pressed fraction0.2; it does not initialize the button beyond the0.5 success threshold.
Factory uses an unlocked-waist HOMIEv2 WBC, a distinct robot USD, a different head-camera extrinsic/FOV, and mobile base commands. Apple geometry/controller/camera must not be substituted.
Official pyproject restricts its uv environment to Linux x86_64. ARM GB300 needs a separately verified runtime; the pinned submodules declare IsaacSim6.0.1 while the model card mentions6.2.
The pretrained expert was task-tuned on900 public demonstrations plus foundation pretraining. New ACT uses810 gradient-training episodes,90 reserved episodes, task-ID conditioning and320×240 RGB. Outcomes compare current systems, not architecture causality.
The90 reserved episodes select ACT checkpoints and calibrate dispersion, so they are training-held-out but not untouched validation. Per-task q95 is empirical score calibration, not failure probability.
Synchronous RPC pauses simulation. Neither measured wall latency nor native50Hz command timing demonstrates asynchronous latency hiding, Jetson performance, or WAN reliability.

Recorded viewer design

Keep every task, seed and failed episode selectable. Display bilateral wrist and hand differences, height/navigation differences, the task-specific uncertainty threshold, actual policy selection and the five-step blend. Missing expert queries remain visible gaps. Label RPC wall latency separately from simulated time.

Factory source revision: 81cf7909b13fe1a961aae9a4e457a5e786e4df8c. Read-only audit of pinned upstream sources; actual runtime physics, camera, model and media validation is still required.