Skip to content
Bud's Wiki

GPC: Large-Scale Generative Pretraining for Transferable Motor Control

TL;DR

Section titled “TL;DR”

GPC is a three-stage framework for training a generative controller over physics-based motor skills that transfers across downstream character-animation and (implicitly) humanoid-control tasks. Stage 1 (Skill Quantization) trains an FSQ motion-tracking policy end-to-end with PPO — discrete latent codes are optimized directly by RL rather than by a separate VQ-VAE reconstruction objective — on a 600-hour motion corpus, hitting a 99.98% tracking success rate. Stage 2 (Generative Controller) fits a GPT-style causal transformer decoder over the resulting FSQ token sequences, which yields robust unconditional generation and emergent human-like perturbation recovery. Stage 3 (CoLA) is a Conditional Low-rank Adaptation PEFT method (DoRA-inspired) that adapts the frozen generative controller to downstream tasks with <1% new parameters, using either RLFT alone or SFT+RLFT for user-designated stylization. A SIGGRAPH 2026 paper (Shi, Jiang, Tessler, Peng), with an interactive real-time demo on Human-Scene Interaction.

Key claims

Section titled “Key claims”
  • End-to-end PPO optimization of an FSQ-quantized motion-tracking policy on a 600-hour corpus achieves 99.98% tracking success, avoiding the learned-explicit-codebook (VQ-VAE) step and its associated codebook-collapse issues [§Skill Quantization].
  • Splitting the motion latent into dd independent scalar dimensions, each quantized to LL fixed levels, yields an implicit codebook of size LdL^d without a separate codebook-learning objective [§Skill Quantization].
  • A GPT-style causal decoder trained on the FSQ token sequences produces robust unconditional skill generation across diverse behaviors, and exhibits emergent human-like recovery under external perturbation without explicit recovery training [§Generative Controller].
  • Nucleus (top-p) sampling at inference restricts autoregressive choices to high-probability tokens, mitigating low-probability outliers while preserving diverse behavior generation [§Generative Controller].
  • CoLA (Conditional LoRA) adapts the frozen pretrained controller to new tasks with <1% additional parameters, via either RLFT-only (regularized by nucleus sampling from the pretrained unconditional model) or SFT+RLFT for stylization when reference motions are available [§Task Adaptation].
  • CoLA-adapted policies preserve the pretrained controller’s robustness and recovery behavior after downstream task adaptation [§Task Adaptation, Preserved Robustness section].
  • The framework composes learned motor skills to handle diverse parkour-style Human-Scene Interaction scenarios without any skill-specific training from scratch [§HSI Task].

Method

Section titled “Method”

The full pipeline is three RL-stage-composed modules over a shared FSQ latent. In Skill Quantization, an encoder–decoder tracking policy maps a sequence of target motion states through a continuous latent, which is FSQ-quantized (each latent dimension independently rounded to one of LL scalar levels) and decoded into per-frame character actions. Crucially, this quantization is trained directly end-to-end with PPO on tracking rewards — no separate VQ-VAE stage, no explicit codebook — which is the mechanism the authors identify for avoiding codebook collapse and getting a robust discrete skill representation.

The Generative Controller is a GPT-style causal-attention transformer decoder whose vocabulary is the FSQ codebook. It models the joint distribution of skill-token sequences conditioned on character state, and generates autoregressively via nucleus sampling. This unconditional model is where the emergent recovery behavior lives.

CoLA (Conditional LoRA) is a DoRA-inspired PEFT method that adds <1% new parameters to the frozen controller. Two adaptation modes are supported: (a) RLFT-only, where task-specific rewards drive fine-tuning while nucleus sampling from the pretrained unconditional model regularizes exploration; (b) SFT+RLFT, where a small set of example motions supervise the model toward user-designated skills first, followed by RLFT for reliable composition. The HSI demo composes learned parkour skills onto varied terrain via CoLA adaptation.

The framework builds on the NVlabs/ProtoMotions GPU-accelerated simulation-and-learning stack (the same stack Chen Tessler was pointing at in the sharing tweet).

Results

Section titled “Results”
  • Tracking: 99.98% success rate on a 600-hour reference-motion dataset with end-to-end PPO-trained FSQ [§Skill Quantization].
  • Comparison: reported to beat other tokenization/tracking methods (specific baselines and numbers deferred to the paper’s tables, not fully materialized on the project page).
  • Generative robustness: unconditional sampling across diverse skills is robust; emergent recovery under external perturbation without explicit recovery training [§Generative Controller].
  • Adaptation efficiency: <1% new parameters via CoLA is sufficient for downstream task adaptation while preserving pretrained skills and recovery behavior [§Task Adaptation].
  • HSI: real-time interactive demo composing learned parkour skills to navigate parkour-style scenes, shown at SIGGRAPH 2026.

Why it’s interesting

Section titled “Why it’s interesting”

GPC sits directly in the middle of two threads already on the wiki. First, it is the third filed motor-control paper this year that treats motion tokenization as a learned latent over which a downstream generative model can compose skills, alongside MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives (VQ-VAE + MaskGIT-style backbone with keyframe primitives) and Scaling Behavior Foundation Model for Humanoid Robots (ScaleBFM — no explicit tokenization, PPO-based motion tracking on 102M frames). GPC’s specific bet — FSQ trained end-to-end with PPO rather than as a separate reconstruction-loss stage — is the sharpest contrast with MotionBricks’s VQ-VAE tokenizer + MaskGIT generator recipe, and it lets GPC skip codebook-collapse mitigations entirely. Second, on the adaptation axis, CoLA is a concrete instance of applying Parameter-Efficient Finetuning to a physics-based motor-control backbone, which was previously conspicuously missing from the concept page (dominated by LLM/VLM LoRA variants). The RLFT-regularized-by-nucleus-sampling-from-the-frozen-prior trick is a nice bridge to the Reasoning RL concept — the pretrained unconditional model plays exactly the role of a KL anchor.

Complements ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting (ReActor’s physics-in-outer-loop retargeting) as the skill-library half of the same problem: ReActor asks how to get a physically feasible reference to track, GPC asks how to compose an already-trained tracker into higher-level policies.

See also

Section titled “See also”