Sergey Levine: why acrobatic tricks are a poor test for a robot (Rubric #Robotics)
In May, I started the robotics series by asking where real engineering progress ends and a well-produced demo begins. Sergey Levine’s interview, published on August 24, 2026, offers a useful test: a spectacular backflip can reveal less about the future of robots than mundane work with an unfamiliar cup in a new room.
Levine is an associate professor of EECS at UC Berkeley and co-founder of Physical Intelligence. His central thesis is that robotics is not yet at the GPT-4/5 stage, where the recipe has been found and the remaining job is to scale the data and model. The field is still determining which combination of architecture, data, and training will scale. In his view, the components already exist, but together they do not yet produce predictable growth. This is a researcher’s perspective — and also the bet of a company building robot foundation models.
The first test is generalization: the ability to transfer a skill to new conditions. A rehearsed trick tests one specific skill; a useful robot must transfer it to a new object, environment, or robot design. According to Levine, in one Physical Intelligence test a model disentangled two shirts it had accidentally picked up and continued folding one — it recovered from the surprise. In another, it failed to open a drawer and began putting utensils into the oven: it did not freeze, but chose a meaningful yet wrong continuation. That is the gap between impressive behavior and reliability.
Transfer depends on the right curriculum. Repeating one welding operation will make a robot better at welding, but not general-purpose: it needs different tasks, environments, objects, and robot designs. Levine also argues against starting with YouTube. Broad real-world robot experience should first give the model physical grounding; human video and simulation can then expand its knowledge. This is a promising research hypothesis, not an established law of the field.
The modality of the intermediate step also matters. Language provides the task sequence: open the drawer, pick up the object, put it away. An image or video shows the next state of the scene — where the arm should move and what should change. According to results from Physical Intelligence, such a visual subgoal helps transfer skills across different robots.
The entire hypothesis, however, runs into the final few percentage points. Even an illustrative 95% success rate means one failure in every twenty attempts. Levine believes that closing this gap will require reinforcement learning from autonomously collected experience. One caveat matters: in the 13-hour experiment described by Levine and the Physical Intelligence paper, a person issued a high-level command roughly every five minutes, including an instruction to clean up after a mistake. It was a long run, but not yet a fully autonomous shift.
Scaling also has an industrial layer. When asked about China, Levine does not frame the answer in terms of winners and losers. His lesson is that models do not take off separately from manufacturing, supply chains, open development, and affordable, reliable hardware. Scaling Physical AI involves the entire system around model training.
Levine’s forecast is cautiously optimistic: structured applications may be arriving now, while settings such as homes could follow within several years, probably in less than a decade. That is a forecast, not a plan. The safety section of the interview is also noticeably thinner than the technical discussion: Levine proposes putting systems into the world, observing what happens, and adjusting the approach. That does not replace testing, constraints, and evidence of readiness, which I recently discussed through the example of Waymo.
This recording is worth watching as a checklist for the next robot demo. Is the object new to the robot? Does it work in another room? How often does an operator intervene? Can the system recover from a failure — and turn it into data? Those are the questions worth keeping in mind the next time a humanoid performs another beautiful backflip.
#Robotics #AI #Engineering #Research #Architecture