Figure's 237 out of 420 became the week's fight
Tony Zhao pulled a 237/420 success rate out of Figure's own blog and said failing half the time is not real useful work. The rest of the field then argued about whether publishing that number was the problem or the point.
Shown this week
Figure's 30-home run ends as a fight over a 237/420 success rate
Figure's four hours of zero-shot work in 30 rental homes ended the week being argued over a single number from its own blog. Tony Zhao cited a 237 out of 420 success rate and said failing half the time is not doing real useful work, defining useful as generalization plus reliability. Keerthana Gopalakrishnan and Chris Paxton both said publishing the rate at all deserves credit.
Let’s not shame people for being transparent about their success rates.
Reward AI's OM-1 zero-shots humanoid and arm demos, footage is 1x speed
Reward AI launched OM-1, a model the company says learned directly from human manipulation data with no teleop and no robot data, and zero-shot generalizes to tabletop arms, industrial arms and humanoids. One demo shows the robot unplugging an Ethernet cable, a task Reward AI calls deceptively hard because the latch has to be pressed precisely.
daaamn, just realised this is on 1x speed
Humanoid Scott says Digit 5 shuffles rather than walks
A week after Agility unveiled Digit 5, Humanoid Scott found a two-second clip of it moving buried in an Agility video about data security, and says what it shows is a shuffling gait, not walking.
Well, at least we know if can move under its own power now.
Out in the world
Watney raises $80M and claims four nines across data center deployments
Watney announced an $80M Series A, over $100M total, and says its systems have cleared more than four nines of reliability across hundreds of thousands of hours in customer facilities. The company says it has served the largest hyperscalers since 2025 and now runs the largest fleet of dexterous robots operating 24/7/365 in the United States. Lukas Ziegler noted Watney was folding laundry in hotels two years ago.
Teleop and other arguments
Jitendra Malik asks whether Astra is plagiarizing the papers it builds on
Jitendra Malik said AI models like Astra should arguably be accused of plagiarism for not citing the research they build on, and by the weekend he and Ken Goldberg were trading jabs about who gets credit. Malik noted he co-authored in-hand rotation papers that may have fed the Astra pen-spinning demo, and compared disrupted research incentives to farmers eating their seed corn.
CV researchers didn't solve robotics. Robotics is still mostly unsolved, though making rapid progress.
Also on the timeline
Chelsea Finn says stacking was harder than building the box
A Pi robot at Dandelion Chocolate ran for hours fully autonomously with no interventions.
RAI's AthenaZero throws a baseball at 113 km/h
Lukas Ziegler wrote up AthenaZero from the RAI Institute, on this month's cover of Science Robotics.
Frequently asked questions
What is Figure's success rate in homes it has never seen before?
Figure reported a 237 out of 420 success rate when operating in 30 rental homes with no new training. However, nobody outside Figure has independently verified this number.
What is Reward AI's OM-1 and how was it trained?
OM-1 is a robot model from Reward AI that the company says learned directly from human manipulation data without teleoperator control (teleop is when a human remotely operates the robot) or robot-specific data. It can perform tasks across different robot types including tabletop arms, industrial arms, and humanoids without additional training.
What reliability level does Watney claim for its robots?
Watney says its systems have achieved four nines of reliability (meaning 99.99% uptime) across hundreds of thousands of hours in customer facilities. These figures come from Watney itself and have not been confirmed by customers or third parties.
Is there proof that Astra plagiarized research papers?
No documented evidence has been shown about what Astra actually drew from, so the plagiarism discussion remains an argument about incentives rather than a confirmed finding.