Owned research — Hyperion
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.
π0.5 full fine-tune — peak memory
68.71 GB
Real fine-tune (LIBERO, language-conditioned) — loss
1.08 → 0.13 (300 steps)
Simulation success (LIBERO-Spatial) — base → fine-tuned
0% → 96%
On-robot cognitive benchmark (12 scenes) — perceive · understand · interpret
12 / 12
bf16 throughput vs FP32 (ACT)
2.3× – 14.3×