Real-time robotics control on NVIDIA Jetson.
A single self-contained runtime, generated end to end by veizik’s own compiler — no CUDA or C toolchain, no external runtime — that runs a robot whole-body control policy as a hard real-time loop on an NVIDIA Jetson AGX Orin.
Install, activate, run.
One self-contained package installs the CLI and all four control policies; one activation per device.
# install (signature + integrity checked on install) $ ./install.sh # license this device $ veizik activate # one-time real-time setup (RT scheduling + pinned clocks) $ sudo sh jetson_rt_setup.sh # run a policy — prints control latency and action fidelity $ veizik run pol-humanoid-sac
Policies in the first set: pol-humanoid-sac, pol-ant-sac, pol-bipedal-td3, pol-hcheetah-td3 (whole-body / locomotion control, 16–32 action DoF). Policy weights are Stable-Baselines3 RL Zoo (MIT); the veizik runtime and precision are veizik’s.
How every figure is taken.
Measured on an NVIDIA Jetson AGX Orin 64 GB (GA10B GPU, 12× Arm Cortex-A78AE), L4T R36.4.3, Linux 5.15 stock (PREEMPT, not real-time), clocks pinned, one CPU core isolated for the control loop.
- Scheduler — the control loop runs on the isolated core under SCHED_FIFO, memory locked, with an absolute-time periodic release and a hybrid wake (sleep to just before the instant, then a short busy-wait to the exact instant) to remove kernel wake-granularity jitter.
- Boundary — we report engine latency: from the expected release instant to action-ready, measured inside the runtime with the monotonic clock. System latency (sensor callback to actuator) is a separate, larger boundary and is never conflated with it.
- Fidelity — every action is compared to the full-precision reference output by cosine similarity, and we publish the number. Precision is veizik compiler-level sub-byte, chosen per model, not a fixed setting.
- Discipline — every number is a measured record (no projections); a tie counts as a loss, and an independent raw-timestamp verifier recomputes the percentiles outside the engine.
Control-loop latency.
Engine latency on the isolated core, in microseconds — zero missed deadlines in every row.
| Scenario | p50 | p99.9 | max | jitter p99.9 | miss |
|---|---|---|---|---|---|
| 50 Hz control (idle) | 21 | 27 | 31 | 0 | 0 |
| 50 Hz + full CPU saturation | 22 | 41 | 45 | ≤3 | 0 |
| 1 kHz torque-rate control | 21 | 24 | 29 | 0 | 0 / 20,000 |
One hour, zero misses.
Sustained hard real-time on a stock kernel with one isolated core.
| 1-hour gate | 50 Hz (180,000 cycles) | 1 kHz (3,600,000 cycles) |
|---|---|---|
| Deadline-miss | 0 / 180,000 | 0 / 3,600,000 |
| Response p99.9 / max | 26 / 45 µs | 21 / 36 µs |
| Jitter max | 1 µs | 1 µs |
| CPU power avg / peak | 2.4 W / 3.2 W | — |
3.6 million control cycles at 1 kHz over a full hour with zero missed deadlines; a raw-timestamp verifier recomputed the miss count and percentiles outside the engine and reproduced it (miss-rate 95 % upper bound ≤ 0.00008 %). Core isolation is required for this tail: without it, rare kernel-interference events produce occasional multi-millisecond stalls.
Action quality.
Each policy’s action compared to its full-precision reference by cosine similarity.
| Policy | action DoF | cosine vs full-precision |
|---|---|---|
| pol-humanoid-sac | 32 | 0.99988 |
| pol-ant-sac | 16 | 0.99978 |
| pol-bipedal-td3 | 16 | 0.99999 |
| pol-hcheetah-td3 | 16 | 0.99998 |
Honest boundary: these are engine-latency figures on an AGX Orin with a stock kernel and one isolated core. Published real-time robotics results from other vendors run on a different, newer device, so we do not present this as a direct head-to-head win; the capability — compiler-native hard real-time with microsecond-class neural control on a stock Jetson — stands on the measured numbers above.