Harvard, Berkeley and NVIDIA Look for a Scaling Law for Robots: +37.4 Points Across Five Benchmarks

A team of researchers from Harvard, UC Berkeley and NVIDIA has published a paper that looks for scaling patterns in robots. According to the Chinese outlet Jiqizhixin, the new approach beats strong baselines by an average of 37.4 points across five benchmarks. Media coverage immediately tagged the work an "o1 moment" for robotics. The metaphor is clear to anyone who followed language models: after a period of "more data and more parameters," LLMs gained a new axis of growth. Robotics now appears to be approaching a similar threshold.

Why Robots Are Cramped by Ordinary Scaling

In recent years, embodied AI has followed a familiar playbook. Researchers collected ever more teleoperation data and built ever larger models, and VLA (vision-language-action) models alongside WAM (world action models) regularly raised the capability ceiling in labs. But real-world deployment runs into a different obstacle. Once the platform or work environment changes, the question arises: how many people, how much data and how many GPU-hours must be invested all over again? Thousands of hours of demonstrations collected for one robot transfer poorly to another.

This is the most expensive part of robotics. A model looks impressive in a demo, but in a warehouse or factory it takes months of fine-tuning. That is why a paper that claims a new way to scale, rather than just another benchmark gain, hits a sore spot for the industry.

What We Know About the Result

Key facts confirmed by the publication:

• authors: Harvard, UC Berkeley and NVIDIA, that is, academia plus a hardware maker for robotics;
• topic: research into scaling laws for robots;
• evaluation: five benchmarks;
• result: an average of +37.4 points over strong baselines;
• comparison context: the class of VLA and WAM models, which are currently considered state of the art in embodied AI.

An average gain of 37.4 points across five tests is unusually large. In robotics, new methods typically add a few percent, and double-digit jumps appear when the problem formulation itself or the allocation of compute changes. The comparison with o1 hints at exactly that: in LLMs, the effect came not from enlarging the training set but from letting the model spend more compute on the problem-solving process itself. For robots, where an error in action costs more than an error in text, such a shift could be especially noticeable.

NVIDIA's Role and the Question of Cost

NVIDIA's involvement is natural. The company is building a dedicated robotics business: simulators, foundation models for humanoids and training infrastructure. If it turns out that robot quality grows predictably as compute increases at a particular stage, that means additional demand for GPUs in both training and inference. When a robot "thinks" longer, the load shifts closer to the device itself or to edge servers.

The downside is obvious. The more compute goes into each decision, the tighter the constraints on latency and power consumption. A robot on an assembly line cannot spend seconds deliberating over every movement. How the authors address this problem will become clear from the full paper and from whether independent groups can reproduce the results.

What Teams Building Robots Should Watch

The practical takeaway for companies deploying VLA models is this: evaluate not only quality on a single platform but also the cost of transferring to another one, meaning how much data and how many GPU-hours it takes to bring task success back to the previous level. If the new work shows that gains are possible without collecting new demonstrations for each robot, teleoperation budgets may be revisited.

It is also worth looking at the benchmarks themselves and checking how close they are to your own scenarios. A double-digit average gain on academic tests does not yet guarantee the same effect on a shop floor. But as a signal that the direction of scaling in robotics has shifted, this paper deserves a careful read.