Twitxy Saved

Simon Mahns @ CoRL

Verified account This account is verified. Learn more

@nautsimon_

Hardware Verification @Axiommathai, 🚗 @drivingbench | prev PyTorch/MTIA @meta, @uchicago
San Francisco nautsimon.com Joined February 2017
1.2K Following 1.6K Followers 410 Posts

Latest posts 5

Pinned post

Simon Mahns @ CoRL Verified account @nautsimon_
Over the weekend, I put an unreleased checkpoint of π0.7 inside an excavator and placed 2nd at the @ActorLabs + @physical_int 36hr hackathon.

With about an hour of embodied data (3 cameras and joystick values, no joint angles), the policy could continuously move dirt with a single prompt.

This was so straightforward to get working (pi is cooking) we spent the rest of the hackathon mounting a Nerf minigun on the roof. We added a people searching/tracking + firing feature to prevent people from getting too close to the dangerous machinery (ai safety feature).

Here’s how we built it + more details (data collection, hardware hax, sim2real, video of me getting shot)🧵
33 replies 25 reposts 291 likes 53K views View on X
Simon Mahns @ CoRL Verified account @nautsimon_
Also a shoutout to @DreamscaleLabs 🏗️. All of our open source wam/vla inferencing and training was done on their platform. They also stopped by while we were running initial tests on the tesla. Cracked team.
DrivingBench Verified account @DrivingBench 1d
Introducing DrivingBench v2 🧵

We wired a Tesla Model 3 so any AI model can drive it, and put 8 frontier VLMs (Claude, GPT, Grok, Gemini models) and 6 fine-tuned robot models (VLAs & WAMs) behind the wheel.

A new lot and two new courses, none of which they’d seen. Only one VLM ever reached the finish, and its fastest run took 26 minutes. 🏁
View quoted post on X
3 replies 2 reposts 18 likes 1.3K views View on X
Simon Mahns @ CoRL Verified account @nautsimon_
I think one of the biggest surprises here was how @GeminiApp's Carbon model in the raw @antigravity harness dominated this task, performing nearly 3x better than the runner up. Not only did it show signs of ICL over turns, but even one / two shotted the course in a few cases, beating a human baseline as well (on course B).

The model was also extremely slow, @tobiges and i sat in the car for hours benchmarking the model, but im sure this will be a solved problem in the next few months.

@GoogleDeepMind is cooking.
DrivingBench Verified account @DrivingBench 1d
Introducing DrivingBench v2 🧵

We wired a Tesla Model 3 so any AI model can drive it, and put 8 frontier VLMs (Claude, GPT, Grok, Gemini models) and 6 fine-tuned robot models (VLAs & WAMs) behind the wheel.

A new lot and two new courses, none of which they’d seen. Only one VLM ever reached the finish, and its fastest run took 26 minutes. 🏁
View quoted post on X