1/ Better robot models come from better data.
That takes two things: hardware that captures it fully, and tooling that turns it into training signal.
We built the first. Today, @GroundedSI launches the second: Grounded API. Access both hardware + API ↓
Incredible day at @solana Breakpoint! ⚡️
Great seeing our R[3]sidency × Construct cohort at the @SuperteamUK booth, catching up with the teams and stopping by our portfolio company @BitRobotNetwork to see their robots in action.
Nothing beats seeing what everyone’s building in person!
Recap #1 from @corl_conf
We shared the keynote stage with @DJiafei on how we’re deploying more robots in the wild to scale robot evals thru @RoboColosseum.
It’s an ambitious mission, but we think it’ll help the industry get clarity on which policies actually work.
First day at Breakpoint 🇬🇧
Won a beanie at the @solanamobile stand, walked the whole expo floor, and tried to Tap to Meet with a @BitRobotNetwork robot.
It reached for my phone. I'm counting that as a tap.
@SolanaEvents, any chance I meet @toly tomorrow?
How @BitRobot Is Solving the Biggest Problem in Physical AI (Interview with @micoolcho)
How do we collect the data robots need—and find out whether their AI actually works in the real world? I met Michael from Bitrobot to explore RoboCap, wrist-mounted cameras, and RoboColiseum, a project aimed at testing robot policies beyond controlled lab environments.
We start with Bitrobot's six-camera RoboCap and discuss how egocentric human video, synchronized wrist cameras, stereo vision, and hand pose tracking can support robot learning. Michael explains the design tradeoffs behind capturing everyday activity and recovering useful 3D information from video.
Then we explore RoboColiseum: independent evaluation across different robot embodiments, fine-tuned and open-world tracks, human judging, and plans for open rollout data. Michael also discusses the project's partners, including Hugging Face, and an announced commitment of 10,000 hours of free real-world evaluation for academia.
We finish with a discussion of simulation and world models, tactile gloves, potential LiDAR integration with RoboSense, and how AI agents could accelerate robotics research. We also look at why network latency matters for robot control and why some physical intelligence needs to run on the edge.
Watch the Full Video Here: youtube.com/watch?v=9Y3RU5u7…
You can benchmark an LLM from a laptop, but robot policies need real hardware / environments for testing.
Yesteday, @pepijn2233 shared how @huggingface is releasing every RoboColosseum eval rollout, including failures.
What a great to wrap up @corl_conf!
10,000 hours of open egocentric data with SOTA 3D head and hand tracking from @GroundedSI are now available 🤗
RoboCap-10k spans 21 industries in 5 continents with 619 real, paid work tasks all using 6-camera/dual IMU RoboCap hardware.
Explore the dataset and research ↓
Excited to announce Robocap-10k. 10,000 hours of ego with metric actions.
This is the high-fidelity dataset I wished I had in grad school but at industrial scale.
It’s time we make robotics model scaling a science.
4/ While 3D labels show how hands move, language adds what they’re doing: one caption per episode, plus timestamped sub-steps.
All 6 raw camera streams ship alongside the labels, so you can re-run better trackers as methods improve.
3/ That calibration is what makes accurate 3D possible. Grounded API from @GroundedSI turns the raw video into metric labels on every frame:
- Gravity-aligned 6-DoF head pose (3 mm median error vs. motion capture)
- Both hands as MANO meshes in world coordinates (4.5 mm joint error)
2/ To keep this much task variety comparable for robot-training, every hour was captured on the same device: our 250g hat with 6 synced global-shutter cameras and 2 IMUs.
Every session is recalibrated to <1 pixel, so there’s no variation between devices across all 10,000 hours.
Check out the hardware in our launch video ↓