Is the GPT of robotics just GPT? Paxton says maybe
Chris Paxton called GPT-6 Astra an incredibly good robotics model, and Chen Tessler replied that it makes zero sense for it to handle short-horizon execution. Meanwhile one lab says Astra ran autonomously for a day and a half and trained its own pen-spinning policy.
Read the whole week in one goFigure's 237 out of 420 became the week's fight
The big one
Astra tests pile up, and researchers split on what it is good for
The RoboDojo evaluation, posted by Tianxing Chen, compares Astra with GPT-5.5 and DeepSeek-Flash in sim and real, plus humanoid high-level control, dexterous piano playing with RoboPianist, and a systematic study of in-context learning. The takeaway is strong semantic and spatial understanding plus impressive in-context adaptation, with physical commonsense as the clear bottleneck. Separately, Andre Infante says he tweaked the harness and hardware setup and got Astra to roughly succeed at the snack tray test on attempt 3, after one run was cancelled for a system prompt issue and another was killed when it dropped the OJ and could not reach it. Not everyone gets that: SkalskiP told Paxton his own zero-shot attempts were mediocre at best, and Paxton answered that most robotics benchmarks are pretty weak.
GPT-6 Astra is an incredibly good robotics model
GPT is both the moment for robotics (in the sense that it shows us we can pretty much solve spatial understanding and reasoning by scaling up), but at the same time it makes zero sense for it to handle short-horizon execution.
At the same time, physical commonsense remains a clear bottleneck
Shown this week
ETH Zurich's robot hand walks around on its own fingers
Lukas Ziegler wrote up an ETH Zurich hand that carries its own weight and crawls using the same fingers it grasps with, no wheels or legs added. Per his thread, the team wrote a reward formulation for the hand's uneven finger geometry and it moves faster in simulation than rewards borrowed from quadruped work, and the hand keeps its existing finger design and position controller. Power and compute are onboard, so it runs untethered, and the listed behaviors include crawling, steering, recovering after being knocked over, typing a sequence of keyboard commands with no vision, and pushing an object to target positions with overhead visual feedback.
The same fingers that grasp things also carry the hand across a surface and hold up its weight.
Grounded ships an API for turning everyday video into robot data
Grounded says its API is live with state of the art results on hand-tracking benchmarks at under 1 cm, state of the art SLAM (simultaneous localization and mapping, figuring out where the camera is while building a map), and in-the-wild egocentric video turned into enriched data in minutes. It integrates with Hugging Face, LeRobot and Rerun, and is built for BitRobot Network's RoboCap suite. Lerrel Pinto said the infra layer for human data is moving, but converting that into an abundance of robots is the unsolved part. Jiafei Duan said RoboCap is the way forward for egocentric data collection.
What’s less clear is how we’ll turn that into an abundance of robots. The next few months should be fun
SOTA on hand-tracking benchmarks (< 1 cm)
Also on the timeline
Astra reportedly trained its own pen-spinning RL policy
@walterzhu8 says he prompted Astra to implement pen spinning with the Sharpa hand in Isaac Lab, build the pen mesh itself, and hand back a trained policy plus a video, and that it ran autonomously for about a day and a half including training. Jason Ma called it a throwback to Eureka, the 2023 work where GPT-4 wrote reward functions in Isaac Gym, which he half-jokingly calls the first self-improvement loop in robotics.
Digit 5's 9-minute recharge gets reverse engineered
Humanoid Scott does the battery math on Digit 5: 90 minutes of runtime and a 9-minute recharge implies roughly a 1.5 kWh pack at 48 V and about 208 A of charge current, which he says points to LFP cells and probably liquid cooling. He labels it speculation, and credits Jordan @LimitingThe for the LFP suggestion.
Tutor unveils Sonny, says it is in customer deployment
Josh Gruenstein introduced Sonny, a human-inspired mobile robot running Tutor's Ti-series VLA (a model that turns camera images and text into robot actions), which he says is in real customer deployment. Chris Paxton flagged the cart-moving as the notable bit.
Cheap hardware day: Innate OS open sourced, Maker Arm ships
Innate open sourced Innate OS, an agentic operating system for general-purpose robots that runs in their simulator or on their sub-$1k open-source robot, with a demo of two robots collaborating. Separately, the first batch of Maker Arm is shipping: a 6-axis arm with 3 pound payload and LeRobot integration at $999 for the DIY kit.