One GPU, three hours: SuperMimic's 1000x humanoid training
Humanoid Scott put Unitree's car-lifting clip under the knife and concluded the robot tipped a balanced shell rather than lifting anything. Meanwhile the cost of training a humanoid to move fell by a factor of a thousand.
The big one
SuperMimic beats 128 GPUs and seven days with one GPU and three hours
A week on, the field has settled on why SuperMimic matters: Qiayuan Liao says it trains a single humanoid full-body policy across 129,514 motions with one GPU in three hours, beating a state of the art that needed 128 GPUs for seven days. Pieter Abbeel's explanation is two lines: a learned world model runs much faster than regular simulation, and backprop through time collects far more signal than policy gradients.
Now it's superfast (and cheap, just one GPU) to train a single policy on *all* the motions.
Shown this week
Mimic runs a video generation model as its robot policy, at Audi
Mimic's FLUX-mimic policy is a video generation model with a thin action decoder bolted on, and Lukas Ziegler says it is running at Audi on kitting, ECU insertion and soft parts like seals and cables.
We've been running a fully autonomous demo of FLUX-mimic nonstop at GTC for two days.
Humanoid Scott says Unitree's H2-X tipped a gutted car, not lifted one
After a week of frame-by-frame work on Unitree's H2-X "Kung Fu Legend Guardian" clip, Humanoid Scott's verdict is that the fulcrum sits slightly forward of the B pillar, near the center of gravity of what he reads as a gutted junker with the engine pulled.
The robot did not lift a car. It tipped a balanced shell.
An agent built a robot sim from a phone video, then the real robot corrected it
RoboSimStudio, from UC Berkeley and NVIDIA Robotics researchers, had Claude and GPT-6 Astra each build a physics twin of a two-arm wiping task from a smartphone video, with no CAD and no calibrated cameras, and one finished in 56 minutes. On the real robot the bar support moved, which the simulation had modelled as fixed, so contact was lost.
Out in the world
Tesla nears 1,000 robotaxis in Texas and publishes a safety page
Sawyer Merritt says Tesla is nearing 1,000 robotaxis registered in Texas, posting what he calls the largest number ever seen together on public roads, and Ashok Elluswamy amplified it. Earlier in the week Elluswamy launched a Robotaxi safety report: 3.3 million autonomous miles to date with zero injuries and zero airbag deployments, against roughly 9 injuries and 4 airbag deployments he says manual driving would expect.
Robotaxi has driven a total of 3.3 million miles autonomously, and has had zero injuries nor airbag deployments
Also on the timeline
Eka raises $150M for a force-first manipulation model
Pulkit Agrawal says Eka's $150M Series A backs a Vision-Force-Action model, led by a16z and F-Prime. His line: force is the language of manipulation.
ETH Zurich has a Unitree G1 climbing ladders
Chris Paxton shared ETH Zurich's G1 climbing and used it to make a point about what demos should show. His take: robots need to be seen interacting with their environment.
Frequently asked questions
What is SuperMimic and why is it significant?
SuperMimic is a training method that trains a humanoid robot policy across over 129,000 motions using one GPU in three hours, compared to the previous approach that needed 128 GPUs for seven days. It works by using a learned world model that runs faster than regular simulation and collects more training signal than traditional methods.
How does Mimic's policy work at Audi?
Mimic uses a video generation model with an action decoder as its robot policy to perform tasks like kitting, ECU insertion, and handling soft parts. When a human takes over the controls to correct the robot, that correction automatically becomes labeled training data to improve the policy.
What did RoboSimStudio do with phone videos?
RoboSimStudio had AI models build a physics simulation of a robot task from just a smartphone video with no CAD or calibrated cameras, completing one simulation in 56 minutes. When the real robot revealed that the simulation had modeled a support incorrectly, that correction went back into the digital twin.
How many Tesla robotaxis are operating and what is their safety record?
Tesla reports nearing 1,000 robotaxis registered in Texas. According to Tesla's safety report, their robotaxis have completed 3.3 million autonomous miles with zero injuries and zero airbag deployments.
What is Eka's funding focused on?
Eka raised $150 million to develop a Vision-Force-Action model that treats force as the primary language for robot manipulation tasks.