Breaking

NVIDIA Unveils GPC, a Pretrained Motor-Control Framework, at SIGGRAPH 2026

June 30, 2026 at 17:27 EDT

Research & Papers
Robotics
Media Generation

NVIDIA Research and Simon Fraser University have introduced Generative Pretrained Controllers (GPC), a framework that pretrains motor control itself at scale and reuses it across new tasks. The work will be presented, along with a real-time interactive demo, at SIGGRAPH 2026 (July 19–23, 2026, in Los Angeles).

SIGGRAPH 2026 · NVIDIA Research

GPT for Motion: Pretraining Motor Control as a Vocabulary of Tokens

Generative Pretrained Controllers (GPC) turn motion skills into discrete tokens and pretrain them GPT-style — reframing physics-based character control from one bespoke controller per task to a single reusable, pretrained motor brain.

600+

hours of diverse motion data pretrained

99.98%

tracking success rate (FSQ-based controller)

<1%

added parameters to adapt to new tasks (CoLA)

Breaking the dataset ceiling

Prior generative controllers (VAE/GAN, VQ-VAE) stalled at tens of hours. Codebook-free Finite Scalar Quantization lets GPC scale far beyond.

~tens

of hours
(prior methods)

→

600+

hours
(GPC / FSQ)

How it works — the autoregressive pipeline

Tokenize

FSQ maps motion skills into discrete tokens

→

Pretrain

Transformer learns via next-token prediction

→

Adapt

CoLA fine-tunes to tasks, keeping pretrained skills

→

Generate

Nucleus sampling drives diverse, interactive control

Emergent, not trained

In real-time physics simulation, GPC produces natural motion plus behaviors it was never explicitly taught — responding to perturbations and recovering from falls.

Physically grounded motion Perturbation response Fall recovery

Where it could go

Reusable, general-purpose controllers for film, games, XR and robotics — tied to NVIDIA's GR00T and Physical AI direction.

Open questions

Needs large datasets and relies on physics simulation; real-world deployment and hardware-specific optimization still await validation.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.