Integration view
On-robot runtime architecture Every process on MB-1 (with the humanoid H-1 addition), grouped by compute target with its rate. Learned components are indigo, classical components slate, certified safety red. The command path runs left to right from policy to drives; the safety path is independent of every learned component.
On-robot runtime architecture Processes grouped by compute target with rates; learned components in indigo, classical in slate, certified safety in red; arrows show the command path from policy to drives and the independent safety stop path. CLOUD / WORKCELL · async ONBOARD GPU · Jetson AGX Thor REAL-TIME CPU · PREEMPT_RT DRIVES & CERTIFIED SAFETY Fleet backend & telemetry metrics 1–10 Hz · bulk at dock Remote assist console on monitor alarms Reasoner API (optional S2) 0.1–1 Hz · e.g. ER-class VLM S2 task reasoner (small VLM) 0.1–2 Hz → subtask, prompt, subgoal Perception for control 15–30 Hz · poses, ESDF map S1 visuomotor policy (VLA / WAM) 5–15 Hz · chunk 50×24 · TRT NVFP4/FP8 Runtime monitor per chunk · OOD, entropy, value stall Humanoid only: whole-body tracker 50 Hz → PD targets · priority stream Action server (RTC, interpolation) 50–200 Hz · clock alignment, blending Safety filter + diff-IK QP 0.5–1 kHz · limits, CBF, workspace Cartesian impedance / joint PD 1 kHz · force caps, stiffness Residual actor (RL) 50 Hz · bounded Δpose State estimation 0.2–1 kHz proprio · 10 Hz LiDAR Logger (MCAP ring buffer) all topics · event upload Drive servo loops 4–40 kHz · current / torque Safety PLC + scanners ≈1–10 ms · PL d / SIL 2-class independent of every learned component: E-stop, SSM, SLS, STO SENSORS · PTP-synced cameras 30 Hz (global shutter, hw trigger) · F/T 1 kHz · tactile 30–60 Hz · joints 0.5–1 kHz (EtherCAT DC) · IMU 0.2–1 kHz · LiDAR 10 Hz · audio 48 kHz STO / stop dashed: humanoid-only path
Sort by rate or compute. The last column is the designed behaviour when the process misses its deadline or fails.
This view renders from the files in data/ with JavaScript.
·· Design rules this layout encodes
No learned component sits on the path between the safety controller and the drives.
Every learned output passes through a classical filter that enforces limits before reaching a controller (stage 09 ).
Slow components (reasoner, policy) are asynchronous to fast ones; their deadline misses degrade capability, not stability.
On humanoids the fastest learned component (whole-body tracker) gets a dedicated priority so a slow VLA chunk can never starve balance control.
Every process emits telemetry with the version tuple; the logger keeps a ring buffer for incident bundles.