Psychedelic AI's Psi-R2.5 flips human demo generation, hits 99% success on assembly

量子位 · wechat · 2026-09-24

Lingchu Intelligence (灵初智能) unveiled Psi-R2.5 with a counterintuitive data-generation trick: instead of converting human demos into robot actions, it runs its Psi-W0 world model backwards—starting from real robot operation footage to synthesize matching human-hand videos. The result is "strongly paired data": scene-aligned, frame-matched, and with robot-side trajectories replayable on real hardware.

The paired data serves two purposes. First, it trains an end-to-end converter so a user can record a phone video of a human task and turn it into robot training data. Second, it enables in-context learning: human demos are converted and fed directly as prompts to the trained model, no retraining required. Psi-R2.5 uses a two-layer architecture—an upper model decomposes long-horizon tasks while a lower model outputs action trajectories from observations.

On data quality, the team cut duplicate task footage, expanded task diversity, and used automated annotation pipelines that decompose tasks into atomic actions. Crucially, they argue pretraining's role is to shrink post-training data needs: for a phone-box assembly task, human-in-the-loop plus RL post-training reached 99% success in 1-2 working days, already deployed at customer sites.

Original post →

Psychedelic AI's Psi-R2.5 flips human demo generation, hits 99% success on assembly

More from Embodied