Robots that learn a task from one demonstration: inside 2026's vision-language-action models

For years, teaching a robot a new task meant hours of hand-coding or hundreds of teleoperated demonstrations. In the second week of November 2026, two of the biggest labs in AI published robot models that claim to need far less.

On 9 November, NVIDIA released GR00T 2. On 10 November, Google DeepMind released Gemini Robotics 2.5. Here's what they do, what they've been tested on, and why it matters even if you never touch a robot.

How a VLA model works, in one paragraph

Take a vision language model that already understands images and text. Train it further on recordings of robots and people doing tasks, paired with the actions taken. Now, instead of answering "what is on the table?", it outputs "move the wrist here, close the fingers this much", many times per second. Larger systems split the job: a reasoning model plans the steps ("pick bracket, place in slot 3, check the tray"), and an action model controls the hands.

Gemini Robotics 2.5: from lab tasks to a car factory

Google DeepMind's announcement evaluated the model on more than 200 tasks, such as serving food with tongs, tying a bow on a gift box and pipetting liquid. The team reports 30% to 100% higher success rates than earlier baselines, depending on the task, with tasks completed 20–30% faster. The same model works on different robot bodies, including Apptronik's Apollo 2 humanoid with a five-finger hand and a two-armed Franka platform with simple grippers.

The headline test was real work. With Apptronik and Mercedes-Benz, Apollo 2 humanoids ran a proof-of-concept on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The task was kitting: picking individual metal brackets from bulk bins and placing them into sequencing trays in the exact order each car needs. A reasoning model (Gemini Robotics ER) tracked the workflow, and the action model controlled arms and hands. DeepMind reports a 96% success rate.

Two details are worth noting:

• Speed of adaptation. Apptronik's CEO says the models improved "dramatically in just a few weeks of post-training" and carried over to the factory with very little drop-off.
• Safety work. DeepMind reports inference-time safety guidance that improved safe task completion around obstacles by 33% and around nearby people by 43%. It also released ASIMOV-Jailbreak, an open benchmark of prompts that disguise unsafe requests to robots.

GR00T 2: learning by watching people

NVIDIA's approach leans on human video. According to The Robot Report, GR00T 2 was trained on diverse robot and human data, including 100,000 hours of NVIDIA's open egocentric (first-person) video. NVIDIA reports that:

• It completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations.
• With one demonstration per task, average success across 50 tasks was 71%. With five, it was 95%.
• It can transfer skills learned from human demonstrations to different robot bodies.
• AI agents can call it through an API to run skills in sequence and recover from errors.

Early users named include Foxconn, applying it to server manufacturing automation, and Siemens, which is evaluating it for industrial use cases. NVIDIA says the code will be available in December 2026.

Why simulation matters so much

Real robots are slow, expensive and breakable. So most training happens in simulation. NVIDIA's Isaac Lab 3.0, released alongside GR00T 2, integrates the open-source Newton physics engine to model deformable objects such as cables and cloth, and reports up to 2.5x higher training throughput than the previous version on a dexterity benchmark.

The hard part is the sim-to-real gap: a skill that works in a perfect simulated world can fail on a real, dusty, badly lit shop floor. That is why benchmarks such as DexBench, which tests generalisation across object, spatial and physics variations, and real pilots like the Mercedes-Benz one matter more than demo videos.

What is still hard

• Reliability at scale. 96% on one kitting task is impressive for a pilot. A production line running thousands of picks a shift needs more, plus a plan for the other 4%.
• Long, messy tasks. Multi-step work in changing environments is where both labs say research continues.
• Safety and certification. Robots near people need physical safety systems and risk assessments, not only smarter models.
• Cost. Humanoid hardware is still expensive compared with a fixed robot arm for a single repetitive job.

What this means for you: skills to learn

You don't need a robotics PhD to work near this field. Physical AI systems are built from parts many software engineers already know:

• Python and PyTorch, and comfort with GPUs.
• Vision basics: cameras, calibration, detection and segmentation.
• Simulation: Isaac Sim or Isaac Lab, and an understanding of why sim-to-real fails.
• Agents and tool calling: high-level planners that call robot skills through APIs look a lot like LLM agents calling tools.
• Evaluation: success rates across many variations, not one lucky video.

The agent and tool-calling side of this stack, planning, tool use, error recovery and evaluation, is a core part of our Applied AI Engineer Bootcamp. Robotics-specific skills such as simulation and control are best learned by building on top of the open models and tools linked below.