Direct weight manipulation beats fine-tuning: pi0.5 hits 75% on cloth-folding
ZeeshanZiaML · x · 2026-11-04
A small experiment on the pi0.5 robot policy's cloth-folding task (results on RoboDojo) compared three approaches: base 50%, fine-tuned on Astra corrections 68%, and direct weight manipulation 75%.
The context is "Weight-Space Programming": researchers had GPT-6 Astra directly edit Qwen3.5-0.8B weights, changing only 0.016% of parameters to fix the model's reasoning loops and improve math by up to 6x and instruction following by 3x — no gradients or training required. A full breakdown is promised soon.
More from Embodied
- Sipeed's 10cm MaixStation Dock claims DGX Spark-class local AI inference, runs off a power bank — vincent_koc · 2026-11-04
- D&A turns any VLM into a zero-shot robot reward model, hitting SOTA on 3 benchmarks — KyleMorgenstein · 2026-11-04
- YC shows robots autonomously microwaving edamame with fine-tuned pi0.5 policies — ycombinator · 2026-11-04
- FLoRe teaches humanoid robots force-reactive loco-manipulation from robot-free human demos — HaozhiQ · 2026-11-04
- Dev builds automated wet lab sim with NVIDIA SIMREADY assets for 24/7 robot experiments — davidfromkansas · 2026-11-04
- Give your robot a Discord account and summon it like a coworker — jaivin · 2026-11-04