πR^2: Enabling 40ms Real-Time Reactivity for Robot Manipulation Policies
CMU-SCS · hf · 2026-07-31
Current large model-based robot manipulation policies typically use open-loop action chunking. Due to high latency in the perception-to-action pipeline, they cannot react quickly to new sensory input mid-execution.
Researchers from CMU introduced πR^2 to achieve real-time reactivity while retaining large backbones. Core innovations include:
- Dual-channel Conditioning: Splits input into a fast channel (proprioception, updated every tick) and a slow channel (vision-language features, asynchronously updated), allowing reactions even with stale vision.
- Latency-adaptive Flow Schedule: Treats in-flight actions as inpainting conditioning, emitting actions in a single denoising step to adapt to varying hardware latencies.
Applied to GR00T-N1.7 on a real xArm6 platform, πR^2 achieves a 25Hz (every 40ms) replanning rate, 4x faster than the base policy. It improved success rates by up to 30% in real-world manipulation tasks.
More from Embodied
- Unitree Turns Spring Festival Gala Performance into UniUni Robot Swarm Choreography Product — CyberRobooo · 2026-10-16
- Tesla FSD (Supervised) approved in Slovakia, the 9th EU country — XFreeze · 2026-10-16
- Berkeley researchers sample humanoid robot policies directly from weight space with diffusion — berkeley_ai · 2026-10-16
- OneWM: unified causal training shows world and action modeling improve each other — IAAI233 · 2026-10-16
- Tesla's hover car isn't a flying car — compressed air can't sustain flight — tinyfool · 2026-10-16
- Scobleizer relays claim that Optimus demo will be even more mind-blowing than the road-flipping Roadster — Scobleizer · 2026-10-16