@eventualai

Data Curation for Physical AI

 San Francisco USA
eventual.ai
Joined September 2022
  • Tweets681
  • Following63
  • Followers900
216Media
Long Horizon Summit: @GenAstronautics builds robots for in-space operations (because space is labor-limited). By founder @bramschork's math, astronauts spent 3,000 hours on maintenance and repair on the ISS last year. That is about $3 billion of labor.
3
2
8
292
more replies
Bram treats space as a data problem. After he saw a robot trained mostly on synthetic data at NVIDIA GTC, he asked which environment has the least data. His answer: space. Next is a parabolic flight later this year. They plan to swap a Jetson GPU avionics payload in microgravity, fully autonomously.
1
2
26
The full interview:
1
3
36
Long Horizon Summit: @TrossenRobotics built their new arms to solve embodiment transfer so that the research arm you train on is same arm you deploy in production. We've got the previous generation in the office and use it all the time to test VLAs/WAMs.
2
9
332
more replies
Their stack is on GitHub: github.com/TrossenRobotics trossen_sdk: records demonstrations to MCAP lerobot_trossen: LeRobot integration trossen_arm_mujoco: MuJoCo models trossen_ai_isaac: Isaac Sim and Isaac Lab

Trossen Robotics

Educational, research, and professional level robotic development platforms. - Trossen Robotics

github.com
1
26
The full interview:
2
14
Long Horizon Summit: @OrangewoodLabs built an industrial arm from sheet metal instead of CNC-machined parts... in two days. 🤯 They also wrote their control system from scratch for VLAs. So we asked why.
2
6
278
Their answer: -VLAs need control that never blocks and sensor data that is fresh and synchronized. -Models output joint actions, joint deltas or Cartesian poses. Most stacks need a rebuild for each. Their controller switches on the fly. -Arm, controller and Jetson: $12,000.
1
48
The extended cut: control loops, singularities, action spaces, and an agent that designs the factory cell.
1
1
88
Long Horizon Summit: @DexmateAI builds robot hardware and the SDKs that run it. So they walked us through it. In the demo, an agent gets primitive skills (segmentation, motion planning, obstacle avoidance) and a prompt. It chains the skills into a full pick-and-place task.
3
8
215
more replies
Of course, daft natively supports MCAP and LeRobot for when you want to work with your logs, but if it's the training signal in the data itself you're looking for, check out our curation demos on our homepage. docs.daft.ai/en/stable/datas…
1
1
31
The full 7-minute interview with @EverettKleven
1
1
4
91
Day one of the @encord_team Long Horizon at Fort Mason The floor was full of teams working on hard problems, so we took a tour and asked them what makes them different. See bio for the latest.
2
10
325
Biggest themes: 1. Specialization is still the name of the game. 2. World Model scaling laws are early. 3. Automating a workflow by 90% isn't enough. The other 10% actually creates more work. Teams closest to a fully deployed workflow are closest to real value.
1
2
43
If you're there for day 2, come find us. go.eventual.ai/IFKdP0h

Get in touch — Eventual

Try Eventual on your data: send a few hours of footage and get back signals for curation and training, from actions and contact to trajectories and quality.

eventual.ai
1
23
Load more