Daniel Ho, head of evaluation at 1X, discusses robot world models, policy learning, and evaluation with the RoboPapers hosts, including Chris.
Ho contrasts video prediction with VLA policies that directly predict action sequences. He says video pretraining can transfer knowledge from human footage to new robot tasks, even when the robot has not seen those tasks in its own training data. In the demonstrations, the model imagines a video from an image and text instruction, then an inverse dynamics model converts the video into actions for the robot.
The training pipeline starts with web-scale video, adds about 900 hours of egocentric footage, then fine-tunes on about 70 hours of robot data. Ho says egocentric mid-training improved results on generalization tasks, including dish scrubbing, though success remained low because the task involves difficult contact and the robot data focused mainly on pick-and-place. More targeted robot data could improve performance, but the study aimed to examine transfer across tasks.
Ho notes that the model can produce useful actions while missing fine contact details, such as where a bagel lands in a toaster slot. He sees robot rollouts, including failures and successes, as a way to improve the model and train on autonomy data. Current inference takes about 10 to 20 seconds to generate roughly five seconds of video, so faster inference and more responsive replanning remain open challenges.
Ho describes action-conditioned world models that predict future video and estimate task value, allowing teams to compare candidate policies offline. Their evaluations showed a useful correlation with real-world checkpoint performance, while acknowledging that hallucinations limit accuracy. He says these models can replay specific deployment failures for targeted evaluation, while classical simulators can provide higher-throughput signals.
Ho argues that humanoid robots can benefit from human video and more natural teleoperation, while acknowledging that data collection may be slower than on other platforms. He expects progress to depend on work across models, evaluation, planning, reinforcement learning, and hardware. He names Veo, Sora, DreamGen, Uni, and MimicVideo as inspirations or relevant work.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Host and robotics analysts Scott Walter, Devang Adhyaru of the Humanoid Hub, Gustav, and Merdad examine 1X Technologies’ Neo hand and its mechanical design, demonstrations, and remaining uncertainties.
Bernt Børnich, CEO of 1X Technologies, explains why the company designed its Neo humanoid for homes and discusses its capabilities, pricing, teleoperation, safety, and path toward autonomy with Wall Street Journal host Joanna Stern.
Bernd, a 1X representative, introduces NEO, the company’s home humanoid robot, and describes its design, household features, and planned path toward greater autonomy.
Peter H. Diamandis talks with 1X Technologies founder and CEO Bernt Børnich and investor Dave Blundin at 1X’s Palo Alto facility about building humanoid robots for homes, learning through real-world use, and the manufacturing and safety challenges ahead.
Bernt Børnich, founder and CEO of 1X Technologies, discusses humanoid robot design, embodied AI, and safety with the RoboBoom host.
Eric Jang, AI team lead at 1X Technologies, describes the company’s humanoid robots and data engine at the Spring 2024 GRASP SFI event hosted by GRASP Lab.
