π0.5 INT8 on a 16GB GPU: 223-step sim task success at ~32ms per inference
lxfater · x · 2026-10-27
A detailed breakdown of running π0.5 INT8 via APXInf on a 16GB RTX 4070 Ti SUPER: an Ubuntu box serves the model while a Mac runs the LIBERO/MuJoCo simulator, communicating over LAN via OpenPI. Every 5 sim steps, dual-camera observations (224×224), arm state, and the task instruction are sent to the model, which returns end-effector deltas and gripper commands. The 'turn on stove and place moka pot' task succeeded in 223 steps with 45 model calls; median inference was 32ms with 71ms WiFi round-trip. The author also covers the out-of-memory and compilation issues hit along the way.
Related event: π0.5 Runs on a Consumer RTX 4070 Ti SUPER(3 posts)→
More from coding & agent
- Void v1.3 ships JSX routes, email aliases, and a wave of Cloudflare fixes — cnakazawa · 2026-10-27
- Dev says building game worlds on Spawn with Opus enables endless shooter maps — TAbrodi · 2026-10-27
- RL environments essay: knowing how to operate software isn't knowing how to achieve goals — ShenRaphael · 2026-10-27
- Rolldown ships Inline Common Chunks to cut small common chunk files — irvinebroque · 2026-10-27
- NUS's DynaFlow Learns Free-Form Multi-Agent Orchestration, Beats AFLOW — NationalUniversityofSingapore · 2026-10-27
- ProAS Fixes Uniform Credit Assignment in Critic-Free Agentic RL — LINs-lab · 2026-10-27