microVLA runs pi0.5 VLA training and inference on a 4GB RTX 3050, 2.5x latency cut
this_is_surabhi · x · 2026-11-16
- microVLA is a minimal pure-PyTorch implementation of pi0.5-style VLA training and inference that runs on an RTX 3050 with just 4GB of VRAM.
- It includes a 10M-parameter pi0.5-style model trained from scratch on LIBERO-Spatial demos (converging in 25 minutes / 8k steps), plus the real pi0.5 (3B) running on the same 4GB GPU in 4-bit, peaking at 3.1–3.2 GiB VRAM.
- Through bfloat16 + CPU offload → NF4 4-bit → trimmed prompts → flow steps cut from 10 to 5, latency per 50-action chunk dropped from 1,662ms to 662ms (2.5x), while LIBERO-Spatial success stayed high (99/100, 98/100, 98/100).
- Both the 10M model and 4-bit pi0.5 complete the same task successfully; code is open source.
More from Embodied
- Koshien robot team shares control stack: nonlinear-opt trajectories hit ±1.2cm average error — 4310sy · 2026-11-16
- Most VLAs lose their VLM's vision-language skills during pretraining — DJiafei · 2026-11-16
- RL environments become a business: Mercor at $20B valuation, Lumos launches long-horizon robot RL envs — 量子位 · 2026-11-16
- Developer's New Robotics BLDC Controller PCB Batch Arrives, Running Unit Tests — wavefnx · 2026-11-16
- YC-backed Ornadyne's flapping-wing drone was classified as a bird by Air Force detectors — import_jmr · 2026-11-16
- Uber starts offering Amazon's Zoox robotaxi rides in Las Vegas — Polymarket · 2026-11-16