PredActor Cuts Diffusion to 5% of Forward Pass for Low-Latency Humanoid Control
chris_j_paxton · x · 2026-11-22
A new research effort, PredActor (Predictive Action Diffusion for Steerable Onboard Humanoid Control), proposes a humanoid control policy that denoises future states and executable actions together from proprioception alone, executing only the actions.
- A transformer interleaves state and action tokens under an asymmetric mask: state queries attend to all state tokens but not action tokens, while action queries attend causally to both streams.
- Predicted states serve as the steering interface, with classifier-free guidance amplifying text-conditioned behavior and classifier guidance pushing predicted future states toward test-time objectives such as a destination or joystick direction.
- Deployed on a Unitree G1 with an onboard Jetson Orin NX.
- Main contribution: the diffusion part accounts for only 5% of the policy forward pass, enabling low latency.
More from Embodied
- Dev builds jepad: a 0.24ms proprioceptive JEPA sidecar for the Microduck robot — victormustar · 2026-11-22
- CoRL2026 community lunch massively oversubscribed at 8x capacity — Lianhuiq · 2026-11-22
- Non-developer builds desk kitten with Codex and M5Stack in 4 hours at OpenAI hardware hackathon — OpenAIDevs · 2026-11-22
- Unitree G1-D humanoid runs on Novares auto line: 6 min of demos, 99.5% pick success — CyberRobooo · 2026-11-22
- A project manager's working review: Pebble Index 01 vs Sandbar Stream AI ring — Scobleizer · 2026-11-21
- OpenAI hardware hackathon in NYC yields voice-chat 3D-printed robot — OpenAIDevs · 2026-11-21