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.

This is measurement-and-feasibility work, not a finished product. Every number below is labelled measured, simulation or dry-run — and what is not yet measured is stated plainly.

What we 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)
Measured
Simulation success (LIBERO-Spatial) — base → fine-tuned
0% → 96%
Simulation
On-robot cognitive benchmark (12 scenes) — perceive · understand · interpret
12 / 12
On-robot, dry-run
bf16 throughput vs FP32 (ACT)
2.3× – 14.3×
Measured

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.