DEV Community

Cover image for NVIDIA just shipped a robot model that learns new tasks from a single demonstration. The economics of programming robots has changed.
xBerry
xBerry

Posted on AI-assisted

NVIDIA just shipped a robot model that learns new tasks from a single demonstration. The economics of programming robots has changed.

At CoRL in Austin, NVIDIA launched GR00T 2 - its new open world action model for humanoid robots. The headline number is 71% average success across 50 tasks when the robot has seen exactly one demonstration. Five demonstrations gets you to 95%. One year ago, teaching a robot a new physical task meant weeks of manual programming or thousands of training examples. That number is now one.

In the same week SoftBank led a $500 million Series F in Agile Robots at an $8.5 billion valuation. It is two weeks after China's regulator blocked SoftBank from acquiring ABB's robotics division for $5.4 billion. Additionally, Boston Dynamics announced that Atlas will begin trials at GlobalFoundries' semiconductor plants in early 2027, its first publicly announced customer outside the automotive industry.

The numbers that are worth remembering:

  • 71% | GR00T 2 average success across 50 tasks from one demonstration per task - rises to 95% with five demonstrations; NVIDIA's own figures, task list not published.
  • 2.5x | Isaac Lab 3.0 training throughput improvement over Isaac Lab 2.3 on vision-based dexterous manipulation benchmarks, using the new Newton physics engine.
  • $8.5B | Agile Robots valuation after $500M Series F led by SoftBank - total invested capital now $2 billion; NVIDIA and Hitachi also participated.
  • Nov 17 | Boston Dynamics and GlobalFoundries announced Atlas trials at chip plants starting early 2027 - first publicly confirmed Atlas customer outside automotive.

What GR00T 2 actually means?

GR00T 2 is built on Cosmos, NVIDIA's world foundation model, and it's explicitly designed to generalize - the same underlying model handles tasks across different robot form factors and environments. The 1-demo result is real, with the caveat that evaluation tasks were drawn from those achievable above 80% in five episodes, so the ceiling was pre-selected. Even so, the trajectory is what matters.

Until now, programming a robot to do something new required either a large team of robotics engineers writing motion controllers from scratch, or thousands of demonstrations to train a task-specific model. Both are expensive and slow. If a foundation model can achieve 71% from one human demonstration and you can collect five demonstrations in under an hour - the bottleneck moves from engineering to data collection. It's a fundamentally different cost structure. Early adopters named in NVIDIA's CoRL coverage include Foxconn, Siemens, Nexuni, and MicroAGI. Those aren't research labs. They're production companies testing whether this translates to real factory floors.

Isaac Lab 3.0 compounds this. The 2.5x training throughput improvement means simulation-based training for dexterous manipulation - the hardest part of humanoid deployment - gets dramatically cheaper. If you can simulate more, faster, the gap between a lab capability and a deployable skill closes. That is the infrastructure layer that makes GR00T 2's 1-demo number meaningful at scale, not just in benchmarks.

GR00T 2 ships with open benchmarks - DexBench and VABAR - developed with RLWRLD and UC Berkeley. Publishing the benchmarks alongside the model is how you invite the research community to either confirm or challenge the numbers. It's a different posture than the closed evaluations most labs have been running.


SoftBank's $500 million bet on Agile Robots

The timing of this is certainly not subtle. On November 5th, China's NDRC blocked SoftBank's $5.4 billion acquisition of ABB's robotics division. And then just a week later - on November 12th - SoftBank announced it led a $500 million Series F in Agile Robots at an $8.5 billion valuation. The message was clear: SoftBank wants a high position in physical AI infrastructure, and if it can't acquire an established industrial robotics platform, it will fund the next generation instead.

Agile Robots is a Munich-based company - it matters. Its production expansion includes Germany and the United States - two markets where the geopolitical calculus around Chinese-owned robotics infrastructure is most acute. NVIDIA and Hitachi Group joined the round. Total invested capital is now $2 billion. The valuation is up from just over $1 billion in SoftBank's 2021 Series C - an 8x step-up in five years. For wider context: SoftBank's Vision Fund has already had a complicated history with robotics bets. The difference this time is that Agile Robots has paying customers and a production pipeline, not just a roadmap.


Briefly: Atlas is heading to chip fabs

Boston Dynamics and GlobalFoundries announced on Nove*mber 17 that Atlas will begin trials at GlobalFoundries' semiconductor manufacturing facilities in early 2027*. It makes GlobalFoundries the first publicly confirmed Atlas customer outside of the automotive industry. The early work is exploratory - workshops, systems integration mapping and task identification before any autonomous operation. Spot will also expand to GlobalFoundries facilities in New York, Vermont, Germany, and Singapore. The significance is less about what Atlas will do in a chip fab in Q1 2027 and more about what industry vertical just opened up. If Atlas works in semiconductor manufacturing, which has extremely high precision requirements and significant labor costs, the addressable market for humanoid robots gets meaningfully larger.

Three different events happened in the last 2 weeks, but we can see one pattern.

The cost of teaching a robot a new skill is falling. The capital going into physical AI infrastructure is accelerating. And the deployments are expanding beyond automotive into industries where the precision bar is higher and the labor-cost argument is even stronger.

The interesting question is not whether humanoid robots will reach semiconductor fabs. It's how long before they are the default answer for why a chip manufacturer does not have enough workers.


Briefing prepared by Klaudia from xBerry - a tech company based in Poland at the intersection of Physical AI and industrial operations.

Top comments (0)