A New Humanoid Robot Model Beat Public Baselines on Just 30 Hours of Practice

A small team at the University of Southern California trained a new humanoid robot foundation model on roughly 30 hours of real, human-controlled practice and beat public baselines by more than 40% on eight long, multi-step manipulation tasks. The training design that produced that result treats human and robot actions as two separate skills, and that decoupling is the point the embodied-AI field will spend the next year arguing about.

The model, called Ψ₀ (Psi-zero), was unveiled this week at RSS 2026 in Sydney, the top-tier Robotics: Science and Systems conference. Ψ₀ lands as a clean challenge to the prevailing view inside embodied AI: that teaching robots to move like humans is mainly a data-collection problem.

Ψ₀'s authors refuse that premise. Their paper and the companion project page argue that human and robot actions are not the same language. Human hands have different joints, different degrees of freedom, and different dynamics from a humanoid robot's grippers and arms. Trying to learn hand trajectories from first-person human video and then transfer them directly to a robot, the team writes, blends two skills that should be kept apart. Transferring visual and semantic understanding is more data-efficient than transferring the motor pattern itself.

The training pipeline runs in three stages. In the first, the model ingests roughly 829 hours of EgoDex, a public dataset of first-person human video captured with Apple Vision Pro, which gives accurate 3D hand positions. The point of this stage is task semantics and egocentric visual representation, not motor control. The second stage freezes the visual backbone, a Qwen3-VL-2B vision-language model from Alibaba, and trains only an action head on top: an MM-DiT-style diffusion transformer borrowed from the image-generator world, where visual and action signals flow through parallel streams that meet in joint attention layers. The third stage is task-specific fine-tuning, again from teleoperation.

The robot data was about 30 hours of teleoperation, with humans controlling the robot arm by arm, used to teach lower-body control and manipulation. That is a constraint of what the lab could collect, not a design claim. The team would have liked more robot data, and the 30-hour figure was a function of available practice, not a hypothesis about minimal data.

On the evaluation, which covers eight long-horizon tasks such as grasp, push, pull, pour, and place, each broken into three to five sub-skills, sampled at 30 Hz, with runs longer than 2,000 steps, Ψ₀ posted a more than 40% absolute lift in success rate over public baselines. The comparison is against selected public baselines, not state of the art. In an ablation, removing the EgoDex pretraining stage drops the success rate to roughly 20%, which is the evidence the authors cite for the design choice: the pretraining is what carries the model across the human-to-robot gap.

The implications run through the field. Today's dominant playbook is to scale real robot hours: collect millions of teleoperation trajectories and let a vision-language-action model absorb them. That is the bet behind most of the major humanoid foundation-model efforts. Ψ₀ puts a different bet on the table: training design, specifically refusing to naively mix human and robot data and learning visual semantics and motor control as separate skills, may matter as much as raw robot hours. The data and code are open-sourced, which means other labs can run the same test with bigger robot-data budgets and see whether scaling erases the design's edge, or whether the design holds.

The honest read is narrower than the wire headline. The eight tasks are lab tasks, not commercial scenarios. The 40% lift is against public baselines the team selected, not against the strongest private systems inside major humanoid labs. The team itself frames the 30-hour figure as a practical constraint, not a claim that 30 hours is optimal. The paper claims, and the ablation supports, that the decoupled design is doing real work. Whether that work generalizes beyond eight tasks and beyond the EgoDex data distribution is the next test the field owes itself.

The next checkpoint is straightforward. Run Ψ₀'s training pipeline with a 10x or 100x larger teleoperation budget, against the same eight tasks, and see if the success-rate gap between the decoupled design and a naively mixed one shrinks or holds. The code is public; the project page is live. The embodied-AI field now has a concrete alternative hypothesis to falsify, not just an argument to have.