Benchmarks and data

This section tracks the datasets and leaderboards that decide how world action models are scored, not a parallel copy of those leaderboards. Numbers on this site always link to the paper or lab page they came from. Author estimates are labelled as estimates.

RoboArena is the distributed real-world comparison Reuss uses as a public Elo snapshot. In the April 2026 figure he publishes, DreamZero is at 1,750 Elo and π0.5 at 1,622. He notes that DreamZero snapshot was trained only on DROID. Reuss, NVIDIA Technical Blog (15 June 2026)

Simulation tables on the register come from the models' own papers. GR-1 reports CALVIN success rate 94.9% against 88.9% in its comparison, and Reuss highlights 3.06 of 5 subtasks on the ABC→D split. Video Prediction Policy reports an 18.6% relative gain on CALVIN ABC-D. Cosmos Policy reports 98.5% average success on LIBERO and 67.1% on RoboCasa. Unified Video Action reports a 20-point gain on PushT Multitask and 5 points on LIBERO-10. Fast-WAM reports LIBERO and RoboTwin results without embodied pretraining. Being-H0.7 reports strongest overall rank across six simulation suites in its own comparison, including LIBERO, CALVIN and RoboCasa.

Latency is a separate score. Fast-WAM Fig. 4, measured on one real towel-folding task on an RTX 5090D, reports 190 ms for Fast-WAM, 580 ms for Fast-WAM-Joint, 810 ms for Fast-WAM-IDM, and 180 ms for π0.5 — about three to four and a half times slower than π0.5 for the imagine-then-execute variants, which is the subject of the speed explainer.

Reuss's field map also says robot benchmarking remains the field's weakest link, and that a full WAM path built on a 14-billion-parameter video model is, on his own rough estimate, around seven times the training compute of an equivalent VLA path. Those are labelled estimates, not paper-reported budgets. Reuss, NVIDIA Technical Blog (15 June 2026)