Dyna's uncut laundry hour hits a dryer-timer question
Dyna Robotics called its hour-long laundry video uncut, and then Steve Crowe asked why the dryer counts down from 32 minutes to 1 in about 30 seconds of real time. That one reply is the week in robotics: the demos got longer, and so did the questions.
Shown this week
Dyna's hour-long laundry run draws a question about the dryer timer
A week after Dyna Robotics released Dyna-2.1 with what it called uncut footage of an entire hour-long laundry workflow, the sharpest reply is still open.
Why does the dryer cycle go from 32 minutes to 1 minute in about 30 seconds of real time?
Agility's Digit ran 12 hours of autonomous tote picking at IROS
Chris Paxton's whole-body mobile manipulation demo on an Agility Robotics Digit ran about 12 hours at IROS, autonomous, at 1x speed, in a location that was not in training data. Jonathan Stephens got the recipe out of Paxton: roughly 500 real-world teleoperated demonstrations for the manipulation policy, a whole-body controller trained in simulation, and 80 to 90 percent success in the booth through crowds and harsh lighting.
Roughly 80–90% success in the booth despite crowds, harsh lighting, and plenty of distractions
Boston Dynamics gave Atlas 13-DOF hands, and the field argued about how they move
Boston Dynamics put 13 degree-of-freedom hands on Atlas, directly actuated rather than tendon driven, and said they were built for high fidelity simulation to enable sim-to-real RL (trained in simulation, run on the real robot). Chris Paxton noted they still carry a 100-plus pound fridge, apparently by making finger joints far larger than a human's and dropping a finger.
Most motions here are impossible to teleoperate today.
Out in the world
RobCo crosses $1 billion, MicroAGI bets everything on post-training
Munich's RobCo crossed a $1 billion valuation nine months after its last round, per Lukas Ziegler, with Sequoia and Lightspeed doubling down and Cherry Ventures and the European Tech Collective joining. RobCo sells modular robot arms as a service to mid-size manufacturers, and says the US is now its fastest-growing market, with production in Austin.
In LLMs, post-training is what turned impressive demos into products. In robotics, that layer barely exists yet.
Teleop and other arguments
Paxton: model-based control has no future, even on wheels
The week's loudest technical fight started with Xiaolong Wang asking Chris Paxton what RL adds compared to doing IK (inverse kinematics, solving joint angles to hit a target). Qiayuan Liao answered that any online optimization for low-level control, IK, ID or MPC, should be abandoned, because an RL policy tracks more precisely and adapts to unknown disturbances, especially on low-cost hardware.
The future clearly is not in model based control, even for wheeled robots
Also on the timeline
NYU's eFlesh is a touch sensor you print for five dollars
The open-source magnetic tactile sensor needs a hobbyist 3D printer, under five dollars of magnets and a magnetometer board. Slip detection generalizes to unseen objects at 95 percent accuracy.
Paxton: robot sports hide how badly robots see while moving
After weeks of soccer and tennis demos, Chris Paxton points at what is missing: external trackers standing in for onboard sensing. Fluid whole-body control, almost no perception during it.
Frequently asked questions
What is Dyna-2.1 and why is the dryer timer question important?
Dyna Robotics released Dyna-2.1 with uncut footage of an entire hour-long laundry workflow that the company says was autonomous. A question about the dryer timer remains unanswered in public sources, so the one-take claim has not been verified.
How did Agility's Digit robot learn to pick totes autonomously?
The robot was trained using roughly 500 real-world teleoperated demonstrations for the manipulation policy, plus a whole-body controller trained in simulation. It achieved 80 to 90 percent success rate during a 12-hour autonomous run at IROS.
Why did Boston Dynamics redesign Atlas's hands?
Boston Dynamics added 13-degree-of-freedom hands (13-DOF means 13 independent joint movements) that are directly actuated to enable high fidelity simulation for training robot policies in simulation that can run on real robots.
What is the difference between RL and IK for robot control?
Reinforcement learning (RL) policies track targets more precisely and adapt to unknown disturbances, especially on low-cost hardware, while inverse kinematics (IK) is an online optimization that solves for the joint angles needed to hit a target. Researchers are debating whether RL should replace online optimization methods like IK entirely.
How much does it cost to make the eFlesh touch sensor?
The open-source magnetic tactile sensor costs under five dollars to make using a hobbyist 3D printer, magnets, and a magnetometer board.