GPC: Large-Scale Generative Pretraining for Transferable Motor Control

Summary

SIGGRAPH 2026 (cite shi2026gpc). Three stages: (1) end-to-end RL trains an FSQ motion-tracking policy that jointly learns a discrete skill vocabulary and a decoder to physics actions, avoiding VQ-VAE codebook collapse; (2) a GPT transformer models grouped FSQ tokens via next-token prediction; (3) CoLA (DoRA+FiLM) PEFT adapts the frozen prior. FSQ tracking 99.98% success / 34.90 mm MPJPE on Bones (~680 h, 343k clips) vs VQ-VAE 99.94/37.92; AMASS 40 h 99.51%. Unconditional sampling yields parkour, recovery, get-up. CoLA <1% new params for steering, trajectory, barrier, platform. Discord post is a HowToPrompt recap that inflates 99.98% as “GPT generates human movement”; the paper’s number is tracking success in sim, not real-robot parkour.

Keywords

gpc · fsq · physics-animation · nvidia · siggraph-2026 · cola · bones · x

Topics

physics-based character animation, motor control, discrete skills

Research notes

• Primary: arxiv abs 2606.29148 (SIGGRAPH 2026, DOI 10.1145/3799902.3811038, CC BY-NC-ND). Discord/X https://x.com/HowToPrompt__/status/2080588702913274184 via fxtwitter (third-party hype). Project https://yi-shi94.github.io/gpc-page/ and https://xbpeng.github.io/projects/GPC/GPC_2026.pdf. Built on NVlabs/ProtoMotions (Apache-2.0). No GPC-specific public weights found. Motion datasets (Bones/AMASS) are prior; not a new hosted corpus, so no datasets_local row.