教机器人干活，光“刷课时”可不够！灵初这次较真数据质量
Ranking
Overall
68
Content
75
Popularity
N/A
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Lingchu Intelligence introduced Psi-R2.5, a robotics model that prioritizes diverse, executable human–robot paired data over simply accumulating training hours. Its data-conversion and in-context learning methods aim to reduce the demonstrations and post-training needed to adapt robots to new tasks.
- Psi-W0 generates aligned human-hand demonstrations from successful robot trajectories, producing strong paired data used to train a converter from human videos to robot imagery and executable actions.
- Converted trajectories are evaluated through real-robot replay and downstream post-training, rather than visual realism alone; some especially complex tasks remain unsolved.
- Psi-R2.5 can use converted demonstration videos as in-context prompts without updating model parameters, while a hierarchical architecture separates long-horizon task decomposition from low-level control.
- For phone-box assembly, Lingchu reports roughly 99% success after 1–2 working days of iterative post-training combining human participation and reinforcement learning.
Sources (1)
教机器人干活，光“刷课时”可不够！灵初这次较真数据质量
Public signals
N/A
TL;DR - Lingchu Intelligence introduced Psi-R2.5, a robotics model that prioritizes diverse, executable human–robot paired data over simply accumulating training hours. Its data-conversion and in-context learning methods aim to reduce the demonstrations and post-training needed to adapt robots to new tasks.
- Psi-W0 generates aligned human-hand demonstrations from successful robot trajectories, producing strong paired data used to train a converter from human videos to robot imagery and executable actions.
- Converted trajectories are evaluated through real-robot replay and downstream post-training, rather than visual realism alone; some especially complex tasks remain unsolved.
- Psi-R2.5 can use converted demonstration videos as in-context prompts without updating model parameters, while a hierarchical architecture separates long-horizon task decomposition from low-level control.
- For phone-box assembly, Lingchu reports roughly 99% success after 1–2 working days of iterative post-training combining human participation and reinforcement learning.