Owned research — Hyperion
Reachy-LeRobot
A feasibility study: running a 4B-parameter vision-language-action policy (π0.5) on a single AMD Strix Halo workstation, with a Reachy Mini robot and SO-101 arms.
What we measured
- Measured
π0.5 full fine-tune — peak memory
68.71 GB
- Measured
Real fine-tune (LIBERO, language-conditioned) — loss
1.08 → 0.13 (300 steps)
- Simulation
Simulation success (LIBERO-Spatial), fine-tuned policy
48/50
Fine-tuned policy: 48/50 episodes (96%). Ten LIBERO-Spatial tasks seen in training, new simulator starting positions, official LeRobot evaluator. One training seed, a separate 1,000-step fine-tune. This is a pilot result on its own; it is not presented as a gain over a baseline.
- On-robot, dry-run
On-robot cognitive benchmark (12 scenes) — perceive · understand · interpret
perceive 10/12 · understand 12/12 · interpret 10/12 · end-to-end 10/12
- Measured
bf16 throughput vs FP32 (ACT)
2.3× – 14.3×
Not yet measured
- On-robot task success rate — the arm loop is dry-run (no physical motion yet).
- NPU (XDNA2) execution — the INT8 toolchain validates, but no DPU subgraph compiled; iGPU/CPU fallback only.
- Convergence (the fine-tune stops at 300 steps), energy, and multi-task generalisation.