End-to-end traces
Two concrete systems traced through all 15 stages. MB-1: a bimanual mobile manipulator mating automotive wire-harness connectors (contact-rich, IL-heavy). H-1: a 29-DoF humanoid doing tote handling and loco-manipulation (RL- and sim-heavy). Each section shows the real artifacts at the hand-off: file formats, tensor shapes, topic names and rates.
The systems in one paragraph each
MB-1 · connector mating
Holonomic base, torso lift, two 7-DoF torque-controlled arms, parallel grippers with visuo-tactile fingertips, wrist F/T, head RGB-D, two wrist cameras, wrist microphone, Livox Mid-360. Jetson AGX Thor + PREEMPT_RT controller + safety PLC. Policy: GR00T N1.7 fine-tune with F/T and tactile tokens, chunks of 50×24 at ≈8 Hz with RTC over 1 kHz impedance control. Acceptance: ≥95 % first-pass success, ≤45 s cycle, ≤0.5 interventions per robot-hour, insertion force <80 N.
H-1 · humanoid tote handling
29-DoF humanoid with dexterous hands. A sim-trained RL whole-body tracking controller (SONIC-class) runs at 50 Hz producing PD targets; a VLA emits motion targets at 10–25 Hz. Data is dominated by mocap, VR whole-body teleop and egocentric human video. Diverges from MB-1 in stages 00, 02, 05, 06, 08, 09 and 14: no statically safe state, sim is the main training ground, and RL owns the fast layer.