/TECH

Guanya Shi proposes hybrid third paradigm for robot learning

DX

GS

It grafts Sim2Real transfer and behavior cloning methods.

Combined views

58.7K

4 posts, first seen 1d ago

431 likes

Reactions from ranked influencers

4 posts

Combined views

58.7K

4 posts, first seen 1d ago

431 likes comments

14 comments

298 saves

58 reposts

Sentiment

Positive86.3%13.7%Negative

Based on 13 visible X reactions from 19 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.

Reactions from the X community

VR vraify@vraify_ai

6:50 AM · Jul 20, 2026

@danfei_xu The 40% definition is the part I’d challenge. A robot that can attempt anything but fails 60% of the time isn’t at a GPT moment. It’s a liability with broad coverage. The real threshold is broad capability plus calibrated refusal: knowing when not to move.

VY Vineeth Yeevani@vineethcyeevani

9:19 AM · Jul 20, 2026

@danfei_xu you should be specific about the embodiments. otherwise the graph makes no sense. i.e what is the difference between DexUMI and ego video?

BY Bob Yong@JustinCarter202

12:32 PM · Jul 20, 2026

@danfei_xu Really impressive pulse check on the robotics industry.

RO Rob@roblee_rl

6:29 AM · Jul 20, 2026

@danfei_xu Great post! There's still a lot of interesting work to be done.

View all

DX Danfei Xu@danfei_xu

5:14 PM · Jul 20, 2026

851TECH

@GuanyaShi yep yep. Incredibly exciting direction.

VN Varun Nair@_varunnair

12:32 PM · Jul 20, 2026

@danfei_xu Great post!

DX Danfei Xu@danfei_xu

6:10 AM · Jul 20, 2026

851TECH

Monday thoughts: Since deep learning, there are two “root-level” paradigm shifts in robot learning: Sim2Real for locomotion and Behavior Cloning (BC) for manipulation. 1/

GS Guanya Shi@GuanyaShi

4:37 PM · Jul 20, 2026

1583TECH

Agreed! One more thing: There is another graft connecting these two trees: BC from human data with simulation-based physics grounding (e.g., retargeting and motion tracking)

14