GPT-6 Astra hits 95% on a robot arm, Malik pushes back
Jay Chooi says a general chat model scored 95% steering a robot arm, Figure committed $3.5B to GPUs for Helix, and Jitendra Malik challenged the LLM crowd to make a legged robot walk.
The big stuff
GPT-6 Astra scores 95% on robot arm control, Chooi says, up from Fable's 40%
Jay Chooi says GPT-6 Astra hit 95% on a robot control task, up from 40% for Fable 5.1, with 6.2x fewer output tokens at 2.3x lower cost. Chooi says the model used no VLA (a model that turns camera images and text into robot actions) and instead wrote out the end-effector pose (where the gripper should go) directly. His earlier Fable 5.1 test was a block into a bowl, 40% up from Fable 5's 5%, on 20 runs each. Separately, lingxiao guo says GPT-6 did real2sim (rebuilding a real robot scene as a physics simulation) from multi-view RGB and robot actions off one prompt: calibrating cameras, building object assets, fitting physics, and running MuJoCo.
GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower cost.
A long way to go before these models can do control tasks, but genuinely real progress here
harnesses + tool calls (IK, SAM3, etc.) can achieve the same result far cheaper, faster, and more reliable.
Deployments and money
Figure commits $3.5B to Nscale for up to 100,000 GPUs to scale Helix
Figure says it is partnering with Nscale to deploy up to 100,000 GPUs on NVIDIA's Vera Rubin platform, committing $3.5 billion up front with plans to go past $6 billion. Adcock says initial deployment starts in the second half of 2027 in Barstow, Texas, and Corey Lynch says the compute is for scaling Helix (Figure's robot AI model). Adcock says there are chapters of the humanoid race money cannot buy, hardware and the AI recipe, and Figure is now in the one it can: scaling intelligence through data and compute.
We're committing $3.5 billion initially, with plans to scale beyond $6 billion
buying a house on mortgage in San Francisco
I’m watching Helix doing things you wouldn’t believe
Research
Huy Ha's Transformer Transformer designs a robot body per task, tested on hardware
Given a target end-effector motion and a reward function, Transformer Transformer generates a whole robot: every link, joint, motor, inertial property, and the controller. It uses RoboTokens (a shared token vocabulary for links, joints, motors, states, and actions) and a diffusion transformer (a model that generates outputs by repeatedly removing noise) trained with different masking schemes, so one network can both design robots and control them. Feed it a cloth unfolding task and it optimizes an ALOHA arm design for lower tracking error and max joint speed; feed it a UMI dataset (demos collected with a handheld gripper) and it outputs a mobile manipulator, where Huy Ha says Sunday Robotics' Memo torso design turned out to be critical.
given target end-effector motion and a reward function, we generate complete robot embodiments for the task, including every link, joint, motor, inertial properties, and the control!
if you run for a single robot on enough tasks, does it recover humanoid?
It seems our janky upside-down mounted aloha arms are optimal all along ;)
Discourse and drama
Jitendra Malik challenges LLM robotics: control a legged robot on varying terrain
Malik's challenge: the Astra demos are simple pick and place with parallel jaw grippers (two-finger pinchers), which mostly shows LLMs can plan, so prompt one to output high frequency control commands for a legged robot on varying terrain, a task he calls five-year-old technology. Phillip Isola said he will see if his group can try, agreeing real dynamic, dexterous control has not been shown and that the best approaches will use high-frequency controllers as tools. Ken Goldberg took a middle position: locomotion can be learned, but he wagers agents using control theory will make more progress on fast, reliable manipulation this year than pure model-free methods (learning control from data with no physics model). Chen Tessler summed up one camp: LLMs should provide the intent, policies should execute, and dense control is a waste of tokens and compute.
The tasks that are demonstrated are simple pick and place tasks with parallel jaw grippers. LLMs can do planning, and the impressive demos in these tasks primarily show that.
I agree that real dynamic and dextrous control hasn't been shown yet. I think it can be done, but I agree that the best approaches will involve high-frequency controllers as additional tools.
People making grand claims about #robotics really need to spend time with actual robots. Debug the controller yourself.
Chris Paxton says Ultra's box-packing video and Agility's Digit couch video are teleop
Phil Trubey posted Ultra footage he called live video from a 3PL (third-party logistics warehouse) customer site: three robots building boxes, packing, taping, labeling, and sorting into USPS and FedEx bins, under a RaaS contract (robots as a service, paid by subscription) with zero payback time. Chris Paxton first called it almost certainly human in the loop, then said his understanding was autonomy with teleop interventions for failures, then said this specific video is teleop (a human controlling it remotely). Agility, meanwhile, posted Digit dragging a couch and picking up chairs as the same generalized manipulation it uses on a warehouse floor. Paxton, calling it our policy, said it is an RL teleop policy where someone is steering it, and still not trivial to pull off.
This particular one is almost certainly human in the loop or just teleop -- that is currently the only way to e2e autonomy
I’m a big fan of Ultra but they also leverage teleop pretty heavily.
This is an rl teleop policy (someone is steering it), but its not trivial to make a robot do such things
Quick hits
Elluswamy calls Tesla Cybercab a real product, Templeton hears 35 running in Austin
Ashok Elluswamy says Tesla's Cybercab is now a real product, open to everybody, with a further step coming next month or so once new tech merges on the v15 plan. Brad Templeton pushed back on fans calling it ready to sell, saying he heard 35 are running in Austin and asking why 100 smooth rides proves the technology works.
X Square Robot's TwinDEX runs a 20-step chemistry task with zero on-robot demos
X Square Robot's TwinDEX pairs a wearable three-finger, nine-DoF hand for data collection with an identical robot hand for deployment, so what a human records is exactly what the robot sees and does. Trained from scratch on a few hundred wearable demonstrations and no on-robot data, it completed a standardized chemistry experiment with tool switches and bimanual work, and the company claims 5.3x the effective data throughput of on-robot teleoperation.
Lightwheel puts 100,000 hours of egocentric video on HuggingFace
Lightwheel released a hundred thousand hours of egocentric video (first-person footage of humans doing tasks) across cleaning, tidying, assembly, construction, and retail, and it is on HuggingFace now. Chris Paxton called it a huge contribution and said the end of the robotics data shortage seems to be in sight.
Amazon has acquired Rivr, Zurich's wheeled-legged doorstep delivery robot startup
Chris Paxton posted a Rivr video with three words, owned by Amazon now, and Lukas Ziegler's Zurich ecosystem map lists Rivr, maker of wheeled-legged robots for last-mile doorstep delivery, as acquired by Amazon. No price or unit numbers appeared in any post we saw.