Skild says $100M ARR, Agility books $1.8M
Agility Robotics' SPAC filing shows $1.8M in revenue against a $140M operating loss, the same week Skild AI said it crossed $100M ARR. Both the money and the benchmarks got a reality check: Lerrel Pinto says academics are in deep despair over big models, and Erik Nieves answered Figure's 1,200 parcels an hour with 2,200.
The big one
Lerrel Pinto sees 'deep despair' as Astra zero-shots robotics benchmarks
Paxton opened it: "If you are a robot foundation model guy and not a deployment guy this must be really concerning." Pinto named Astra, Fable and Muse as models zero-shotting benchmarks in robotics and world models, and called it an uneasy pill to swallow. Goldberg replied that it is also thrilling to witness a paradigm shift; Black told Pinto the field has been here before and to treat it as a chance to jump ahead. Black's longer post goes further, arguing that papers at a conference are often two years out of date and that every paper should open with a hard attempt to solve the problem using existing large models plus a report on where they fail, and Jeannette Bohg agreed, saying conferences should go all in on workshops and panels.
If you are a robot foundation model guy and not a deployment guy this must be really concerning
I’m sensing deep despair in academics over the past week. Astra, Fable, Muse are zero-shotting benchmarks in robotics & world models.
most of the robotics baseline are not so good -- theyre very contact light
Reward AI says OM-1 trained on human data only, no teleop, no robot data
Reward AI introduced OM-1, a robot foundation model the company says zero-shot generalizes to tabletop arms, industrial arms and humanoids after learning from human manipulation data, with no teleoperation (a human driving the robot remotely) and no robot data at all. Yunfan Jiang flagged the number to watch, a new task from under 30 minutes of data, and said keeping each sensor at its native sampling rate instead of resampling everything to a common clock looks like a big reason it runs at human speed. Lukas Ziegler noticed the video is at 1x, not sped up. Animesh Garg called the speed and generality impressive, and told @KyleMorgenstein he caught the same error in the video, which neither post describes.
zero-shot generalizing to any robot: table-top arms, industrial arms and humanoids
I saw the system in action at @RewardAI_ and could hardly believe what it was capable of.
New task from <30 min of data is the number I'd watch.
Out in the world
Skild says $100M ARR; Agility's SPAC filing shows $1.8M revenue
Skild AI said it crossed $100M ARR within 10 months of starting deployments, on factory lines, construction sites, in kitchens and in data centers. Humanoid Scott said the figure is annual recurring revenue on multi-year contracts and that Pathak verified it with him face to face. Lukas Ziegler's roundup adds 60+ paying customers, 90% of revenue from manipulation, and named work with Foxconn, Sumitomo and Mitsui. Kat Scott, meanwhile, posted Agility Robotics' SPAC numbers: $1.8M in revenue on a $140M operating loss, with Steve Crowe saying Agility needs its $300M customer order to come sooner than later.
We just hit 100M ARR within 10 months of starting deployments.
@deepakpathak personally verified the current ARR numbers with us and added this clarification to his earlier article
$1.8M in revenue on $140M operating loss.
Teleop and other arguments
Erik Nieves answers Figure's 1,200 parcels an hour with 2,200
Nieves has 35 years of deploying industrial robots behind him, 200+ deployments and two billion picks. His stream ran 2,200 to 2,400 parcels per hour, bulk flow, full SKU mix, against the 1,200 per hour on Figure's stream. Porter's line in the writeup: "The gap between a demo and an entitled workload is the whole conversation."
The gap between a demo and an entitled workload is the whole conversation.
Astra demos draw reproducibility pushback from Paxton, Tessler and Goldberg
Paxton: "everyone still astra-posting and yet i seem to keep finding problems it can't solve." Tessler said the posts look amazing but he has no idea how to reproduce them, and Goldberg asked why authors with no previous publications posted a potentially major result anonymously. Jiafei Duan asked a poster the plainer question: "How much does it cost for this single run?" Paxton is not all skeptic, he called a commercially available model doing this well still pretty impressive, and he also reposted his own guide to faking a robotics paper ahead of the ICRA deadline: never show failures, pick weak baselines, control the demo environment, do not let outsiders run comparisons.
everyone still astra-posting and yet i seem to keep finding problems it can't solve. curious
They look amazing, but I have no idea how to reproduce them.
This is potentially major news but the authors have no previous publications…Why is it anonymous?
Also on the timeline
Unitree open-sources UnifoLM-WLA-1.0 humanoid foundation model
Unitree fully open-sourced UnifoLM-WLA-1.0, an embodied foundation model it says reaches new state of the art across multiple benchmarks among open-source models worldwide. One model covers desktop and whole-body mobile manipulation, with cross-task and cross-end-effector generalization, per the company.
Maven Robotics leaves stealth with $100M for warehouse palletizing robots
Maven Robotics came out of stealth with a $100M Series A led by RoboStrategy, building wheeled dual-arm robots that move at 10 mph and lift 30 kg for mixed palletizing in warehouses. CEO Hamza Derbas on bipedal humanoids for that job: "They make zero sense for anything they're doing. Very complex, unreliable, and add unnecessary cost." Lukas Ziegler's writeup says the plan is 250 third-generation robots with a fourth-generation platform in design.
Chris Paxton says humanoid factory robots are kitting while humans do the hard parts
Paxton watched footage of a new factory for humanoid robot production and pushed back on the robots-building-robots framing. His read: the robots are kitting (gathering the parts for an assembly) and carrying things to human workers, and that gap is how you know we are not heading for a hard takeoff yet.
ETH Zurich humanoid jumps to monkey bars and swings across
A humanoid from ETH Zurich's legged robotics lab jumps up to a bar structure, brachiates across it (swings hand over hand) and drops to a landing. It never builds a terrain map, because a heightfield throws away thin overhanging geometry like a bar with air above and below it, so the policy reads the raw scan from a head-mounted solid-state lidar with no reconstruction step in between. Training was a phase-scheduled teacher-student setup in simulation that also modeled battery voltage sag and actuator thermal limits, per Lukas Ziegler's breakdown.