signal量子位2026-09-25
Teaching Robots to Work: Clocking Hours Isn't Enough, Lingchu Takes Data Quality Seriously
Lingchu Intelligence's Psi-R2.5 model improves data quality by reverse-generating human demos from robot data, creating strong paired data. It enables in-context learning where a video demo can guide the robot without retraining. In a phone box assembly task, post-training reached about 99% success within 1-2 working days.
- for who
- Robot learning researchers and deployment engineers
- why now
- Figure's Helix 2.5 and Linchu's Psi-R2.5 launch this month, making human-robot data alignment crucial now.
- what changes
- They can obtain high-quality paired data without manual human-robot alignment, and use video demos as prompts for new tasks.
- to do
- Read the Tech blog and apply the reverse data generation and ICL methods to their own robot tasks.
key points
- Psi-R2.5 reverse-generates human data from robot data for strong pairing
- Video demos serve as context prompts, no retraining needed
- Phone box assembly: 99% success in 1-2 working days
#robot learning#data quality#Lingchu Intelligence
score
score 7 out of 10. 0-10: how dense the facts are, multiplied by how much you can do with them after reading. 8+ means the topic's evidence bar is met: benchmarks and availability for a new model, amount and investors for a funding round, revenue figures for a solo-money story. Below 5 an item does not enter the digest. A press release scores 3 or less, a reprint loses 2, anything older than 14 days loses 1, a headline that misleads loses 3.
read the source