Figure, NVIDIA, Physical Intelligence and Google DeepMind publish results for robot models that do different work on different hardware. Their latest releases do not provide a shared test across all four. Putting the headline percentages into a table would give readers a ranking the experiments cannot support.

The current models

Figure: Helix 2.5. Announced September 17, 2026, this model controls Figure’s humanoids. The company’s main evaluation covers tidying living rooms, folding towels and making beds in 30 previously unseen homes. It measures how training on human behavior helps familiar tasks transfer to new places.

NVIDIA: GR00T N1.7. The April 17, 2026 early-access release is a 3-billion-parameter vision-language-action model. A VLA takes images and instructions and produces robot actions. NVIDIA publishes the weights and code under commercial licensing terms, with tools to fine-tune the model for other robot bodies.

Physical Intelligence: π0.7. Released April 16, 2026, π0.7 replaces the original π0 as the company’s current named generalist research model. Its instructions can include language coaching and images of an intermediate goal. The company tests how the model combines learned skills and transfers them across tasks and robots.

Google DeepMind: Gemini Robotics 2. The July 30, 2026 announcement introduced whole-body control and separated the family into a VLA, an embodied-reasoning model called ER 2, and an On-Device 2 model. A planner that reasons about a task and a policy that produces motor actions do different jobs. They should not occupy interchangeable rows in a benchmark.

There is also a newer announced entrant: Agibot’s GO-3 Preview, dated October 29. The company plans to open-source it by year-end. Until that happens, outside developers cannot treat its reported benchmark leads as results they have reproduced.

Read the scoring rule before the percentage

Figure reports 56 percent full-task success for Helix 2.5 with Index pretraining, against 9 percent without it. The experiment held the task data, architecture, training and evaluation fixed. “We give no partial credit,” Figure writes. Every item had to be tidied or every towel folded for a trial to count.

That is useful evidence about pretraining in this experiment. It does not mean the robot can do 56 percent of household jobs. Nor does “zero-shot” mean it learned those tasks without examples: the tasks were adapted using data collected elsewhere, while the evaluation homes and objects were held out.

Physical Intelligence’s π0.7 report asks a different question. Its appliance experiments show a robot improving when given step-by-step language coaching. That is a test of how well the system follows new instructions. A coached trial cannot be treated as equivalent to a fully autonomous run without disclosing the help.

What an outside comparison needs

Start with the same job, hardware and time limit. State what data each model received, keep the test objects out of training, and count failed attempts and human interventions. A robot that finishes after several resets should not receive the same score as one that completes the job unaided.

Access is another practical difference. GR00T’s published weights allow outside testing. DeepMind’s VLA remains in private preview. Physical Intelligence opened its Pi API beta on October 13, letting participants upload data, fine-tune models and use hosted inference. That service is a route to testing on their own robots, not an unrestricted release of every research model.