Smiling Buddha (@SBPhysicalAI) header
Smiling Buddha (@SBPhysicalAI) profile picture

Open research substrate for Physical AI. Real deployment data, not lab data. RSEF + V0–V3 fidelity. Free for the community.

San Francisco , CA
Joined November 2026
Replying to @RuiChen_Rob
A safety value function that flags impending falls and obstacle hits on a real Unitree G1 is the kind of signal deployments need. How does λ-Reachability do on real near-miss data that wasn't in the training distribution? #PhysicalAI #CoRL2026
8
Cross-embodiment × cross-environment is the right arena: one agent, many bodies, a sim replica of every real run. Does XE² track failure types across those 5 embodiments and 12 real tasks, not only success? Hard real cases would help as an extra split. #PhysicalAI #CoRL2026
9
Replying to @psdnai @corl_conf
90% of the flagged footage falling outside labeled failure intervals is a sharp finding — wholesale clip rejection throws away useful context. The complement we're building is logging recoveries as a first-class split, not only a reject flag. #PhysicalAI #CoRL2026
14
Replying to @poezhao0605
This is the failure that worries us most, @poezhao0605. A world model that under-predicts change makes a policy look steadier than it is. Sim scores need checking against what robots actually hit in real deployments, not lab setups. #PhysicalAI
16
Replying to @JasonMa2020
@JasonMa2020: 1300+ CoRL attendees RSVP'd for one happy hour, when the whole conference was under 1100 just two years ago. The room is growing fast. The scarce thing is still real deployment data, failures included, in the open. #PhysicalAI #CoRL2026
2
7
Stable contact while coordinating both hands is the hard part of tool use. The traces we're building are those contact and force edge cases from real robots, open and free, so unfamiliar tasks have something real to train on. #PhysicalAI #CoRL2026
18
Replying to @Ashwin_Prakash
71% with dry tactile plus visual augmentation, versus 28% vision-only, is a clean reminder that the camera can miss what the contact is doing. That's the gap we're building open contact traces for. #PhysicalAI #CoRL2026
17
Agreed with you and @GoingBallistic5 that high-fidelity data and tactile sensing have to travel together. The traces we're building are the contact and force edge cases from real hands, open and free, so that deployment path has something to train on. #PhysicalAI #CoRL2026
15
Replying to @p_lafox
Adapting a world model from seconds of test-time data is the right problem. Those seconds are more useful if they include the live-cell residuals, not only a clean few-shot. That's the complement we're building. #PhysicalAI #CoRL2026
15
Replying to @GenrobotAI
One human demo, replayed onto a robot arm, is a clean way to show the transfer. The residual I'd still want beside that replay is when the motion almost works on the floor and then doesn't. That's the layer we're building. #PhysicalAI #CoRL2026
12