Video: youtube.com/watch?v=4whgw2gL… Blog: bostondynamics.com/blog/robo… (5/5)

Our hand was designed exactly for this. The actuation and backdrivable transmission make it fast and accurate to simulate, which makes RL far more effective at training robust, dexterous policies. Being easy to simulate was a hardware requirement, not an afterthought. (4/5)

Hands need a reactive low-level policy just like the whole-body policy that balances Atlas. Instead of balancing the robot, this policy balances the object. It regulates contact forces, reacts to slip, and keeps the object stable. (3/5)

Such dexterous tasks are near impossible to teleoperate, due to the lack of tactile feedback and the embodiment gap. Visually checking that each finger is in contact while also observing the scene makes teleoperation of multi-fingered hands slow and error-prone. (2/5)

How should we control multi-fingered robot hands? 🖐️🤖 With RL policies. Using policies trained in simulation, our new 13 DoF hand for Atlas can autonomously rotate a ball between its fingers, spin a drumstick, and thread a nut. 🧵 (1/5)

Technically, there is a fourth 🤖 in the background

Video: youtube.com/watch?v=ezPLXk78… (5/5)

Most people love talking about AI, but it's only part of what keeps the robot standing. The unglamorous stuff becomes critical when you don't get a second chance. The hardware, firmware, and communication have to work flawlessly when it counts at the World Cup! (4/5)

Wireless communication gets hard in a stadium where thousands of devices compete for bandwidth. A few dropped heartbeat packets trigger the e-stop, and the robot collapses on international TV. (3/5)

You can't deflate the ball for better grip when it's the same ball Haaland and Neymar use to score the next goal. ⚽️ (2/5)

Tricks that work in the lab don't work at the World Cup! 🧵(1/5)

A few years ago we wondered whether the WAMs could juggle 5 balls. Turns out, they can! 🤹 The hardest part often never makes it into the paper. The system integration behind this is what's truly impressive. Great work @KaiPloeger & @Jan_R_Peters !

If you want to work with us and @_albertorod_ at @BostonDynamics, we are also hiring Research Scientists and Research Engineers RL: bostondynamics.wd1.myworkday… BC: bostondynamics.wd1.myworkday…

More details are in our blog post: bostondynamics.com/blog/can-…

With Atlas, it is now easy to leverage DeepMimic and domain randomization to create dynamic movements ⚽️. For the Ghost Rabona, we first demonstrated the trick via motion capture, retargeted the motion to Atlas, and learned a policy to track the reference in simulation.

Therefore, DeepMimic is one of the most influential humanoid robotics papers of the last decade. Back then, we had many of the ingredients, but were missing the hardware 🤖.

In 2018, I didn’t recognize the importance of the original DeepMimic paper. At the time, humanoid robots were rare and domain randomization was still relatively new. Today, many demos and whole-body controllers are powered by DeepMimic-inspired methods. 🧵

If you also want to work with us and @_albertorod_ , we're also hiring Research Scientists and Research Engineers: RL: bostondynamics.wd1.myworkday… BC: bostondynamics.wd1.myworkday…

Everyone's talking about VLAs, the problems with teleoperation, using sim-to-real RL for dexterous manipulation, and whether tactile sensing is actually necessary. If you're working on any of these, reach out. I'm at ICRA all week and happy to talk.

Day one of ICRA 2026 done! 🤖 Already feels like the biggest robotics conference I've been to. The expo floor is packed with new anthropomorphic hands, half-humanoids and mini humanoids.