How Humans Are Training the Next Generation of Robots: Matt Trossen on The Robotics Stack Podcast
Robots that learn need people to teach them. How that teaching happens, and what hardware makes it work, was at the center of our founder Matt Trossen's conversation with Léo Kharon and Oskar Telker on episode 21 of The Robotics Stack podcast.
They covered a lot of ground: how a company that started selling USB components in 2004 ended up powering the ALOHA project, the new product lineup, and where physical AI is likely to work first. Watch the full episode below, then read on for four ideas worth taking away.
1. Bad data in, bad data out
Most people look to the model when robot learning struggles. Matt pointed at the hardware.
When a full robot arm is used as the leader arm in teleoperation, it needs gravity compensation to hold itself up. The arm can't tell the difference between gravity and the person guiding it, so slow, careful movements turn into a fight with the servos. Operators overshoot, knock objects over, and the demonstrations the robot learns from get worse.
Matt watched exactly that happen during demos at ICRA two years ago and decided it had to be fixed. The team tried smaller motors, springs and belt drives before starting over completely. The result is Glide, a lightweight, encoder-based leader arm with spring-and-cam gravity compensation built into the joints. With no servos fighting the operator, movements stay smooth and the data stays clean.
Glide works with other companies' arms too, so a team with an existing fleet can turn its old leader arms into followers.
2. Training data will come from a mix of sources
Language models had the whole internet to learn from. Robots need physical-world data, and most of it doesn't exist yet. Teams are collecting it three ways:
Teleoperation: a person drives real robot arms, which captures how the arm actually moves.
Egocentric video: cameras worn on the head record a person doing the task with their own hands.
UMI-style grippers: a person does the task with a handheld gripper like our TRumi, which records what the gripper sees and how it moves.
Matt's view is that no single method wins. Egocentric video is cheap and easy to scale, but researchers tell us it lacks the arm's movement data. A blend may get a model most of the way, with teleoperated demonstrations covering the last stretch. That's why we build hardware for all three.
3. The deployments that work start narrow
The teams closest to real returns aren't chasing a robot that can do everything. They pick one task, stick to one robot, and collect data around that task until it works. Matt's second test: can the job be done with a simple parallel gripper, without sensors in the fingers? If so, it's a strong candidate.
Where does he expect physical AI to succeed first? Work traditional automation can't handle and people still do by hand:
Anything involved in feeding people, from processing food to washing dishes
Machine tending, unloading one part and loading the next
Soft, floppy materials like textiles, such as packing shoes into a box with tissue paper and tucked laces
Physical AI won't replace traditional robots that already do precise, repetitive work well. It takes on the messy, variable layer they can't.
4. Hardware is only half the job
Matt describes Trossen as a picks-and-shovels company for the people prospecting in physical AI. Early teams built everything themselves: SDKs, data pipelines, storage, labeling tools. Most teams coming in now would rather buy that ready to go and focus on their application.
So the goal is a robot you can take out of the box, connect to Trossen Cloud and start collecting data with the same day, then train, run the model back on the robot and move toward deployment. Trossen Cloud is in alpha now.
The new lineup Matt walked through, all built in Chicago:
WidowX Arm Series: three arms for different budgets and payloads, from education up to the Heavy's 1.4 m reach and 6 kg payload.
Workbench: two forward-facing arms, compute and cameras on one portable station.
Glide and Cockpit: the lightweight leader arms and the operator station for teleoperation.
Rivet: a mobile manipulator with a vertical lift to reach shelving.
TRumi: our handheld UMI-style gripper for collecting demonstrations without a robot.
Listen to the full episode
The conversation also covers the ALOHA project, why designs across the industry are converging, humanoids, and what it takes to scale a robotics company. Thanks to Léo and Oskar for having Matt on.
Explore our new hardware at trossenrobotics.com.
Comments