Psi-R2.5: PsiBot's robot model learns new tasks from one human demo in 1-2 days
机器之心 · wechat · 2026-09-23
PsiBot released Psi-R2.5, a new embodied foundation model focused on turning human data into usable robot training data.
- Pair Data pipeline: strong pair data is built by reverse-generating human-hand data from real robot data (frame-aligned, replayable actions), then an end-to-end human-to-robot conversion model lets a single phone video become training data for different robots.
- Two-tier architecture: Qwen3.5-4B for long-horizon task decomposition, Wan2.2-I2V-5B for action trajectory generation.
- HIL + RL post-training: fine assembly tasks (phone-box assembly) reach 99% success within 1-2 working days.
- In-Context Learning: one human demo serves as a prompt-like context for zero-shot generalization to new tasks, no parameter updates needed.
- After reaching 100k hours of human data, the team now prioritizes data density over raw scale, with a 50-task real-robot eval set.
More from Embodied
- Steering VLA robot models at inference time lifts grasp rate from 14% to 74% without retraining — drmapavone · 2026-09-23
- Frontier AI models can control robots to follow harmful requests, NBC News reports — ycombinator · 2026-09-23
- 'Robot-use' works on humanoids: coding agents zero-shot control a full humanoid for pick-and-place — ai · 2026-09-23
- REAL: an embodied agent that explores real rooms and asks clarifying questions, accepted at ECCV 2026 — jiqizhixin · 2026-09-23
- Why a humanoid robot auto-damps in locked standing: stiff joints shatter on falls — carlosdponx · 2026-09-23
- An AI-powered BD company to help robotics firms find high-value use cases — gan_chuang · 2026-09-23