MARS Policy: Zero-Latency Multimodal Robot Actions via Adaptive Noise Injection

chris_j_paxton · x · 2026-08-06

MARS Policy introduces a novel generative policy for robots that combines the best of both worlds: the multimodal expressiveness of diffusion policies and the inference speed of deterministic regressors.

Current generative policies capture multimodal action distributions but suffer from high latency due to denoising loops. Deterministic policies are fast but suffer from mode-averaging when multiple valid paths exist. MARS tackles this by adaptively controlling the noise level in action-to-action flow matching.

It injects noise only when multimodality is needed (e.g., at path branching points) and reduces it to zero for deterministic trajectories. Experiments on simulated and real-world manipulation tasks show that MARS achieves near-zero latency and improved success rates compared to standard flow-matching.

Original post →

More from Embodied

Embodied channel →