VibeusGPT-6 Astra leads Playground’s first Robot Intelligence Index,... - Vibeus
GPT-6 Astra leads Playground’s first Robot Intelligence Index, which combines 16 public robotics benchmarks rather than picking a single favourable test. In the default overall view, covering 30 models with both simulation and real-robot results, Astra scores 1905, followed by InternW0-Δ at 1828, GPT-6.1 Sol at 1808, Gemini 4 Argon at 1797 and MWA-Phoenix at 1774. These are fitted comparison ratings, not task-success percentages.
The winner changes with the view. GO-3 Nano Preview leads simulation and tops three individual benchmarks. Dream-1 has the highest real-robot rating from its RoboDojo-RC result; Astra leads the twelve models with results on at least two real-robot benchmarks. OpenWAM-α leads pick and place, Sol household tasks, Astra precision and Opus 5.5 robustness to changes in the scene or instructions. The fourth attached chart is the pick-and-place view, where OpenWAM-α is first, rather than the overall ranking.
Coding agents and world action models occupy the overall top ten; the highest VLA is MolmoAct2 at twelfth. Some coding-agent evaluations involve writing code to control an arm without robot-specific training. That does not make all the underlying tests equivalent: they use different robots, tools, task sets and agent harnesses. Some robot policies were fine-tuned separately for each benchmark, yet matched as one model by name and version. The index therefore combines the model with the way it was prepared and run.
Cross-benchmark disagreement is substantial. Ten of the nineteen models that finish in a benchmark’s top three land in the bottom half somewhere else. DM0 is second of 22 on RoboChallenge but thirteenth of 13 on RoboDojo’s real robots; OpenVLA-OFT is second of 11 on LIBERO-Plus but sixty-third of 63 on RoboDojo simulation. Preparation differs too: RoboChallenge entries can be task-specific models with up to roughly 1,000 demonstrations per task, while RoboDojo fine-tuned DM0 and Spirit v1.5 on 100 demonstrations per task. A high finish on one suite does not transfer automatically to another.
Playground turns within-benchmark score differences into noise-aware pairwise comparisons, then fits a weighted Bradley–Terry model on a chess-style scale centred at 1500. Evaluator independence and repeatability determine benchmark weights; self-reported paper results receive less weight, as does AA-Robotics’s single attempt per task. The public results were collected as of October 28, 2026. The full set contains 68 matched models; the default overall ranking requires both real and simulated evidence, so models tested only in simulation are absent from that view rather than proven weaker.
The displayed rank ranges cover 90% of 300 re-runs that resample benchmarks and perturb scores. Astra’s overall range is 1–5, InternW0-Δ’s 1–6 and Dream-1’s 1–20. Real-robot comparisons are especially thin: only GR00T N1.7 and MolmoAct2 connect RoboDojo-RC to the other real-robot boards. Those connections disappear in about a third of real-robot resamples, making that view’s uncertainty less reliable. Task choices matter as well: RoboDojo-RC selected tasks after trying Astra, and half of RoboColiseum’s score is instruction following and spatial reasoning.
In Playground’s robustness checks, equal benchmark weights and leaving out each benchmark in turn preserve Astra’s overall first place. Extreme random reweighting can change it, and places two through six depend more on the assumptions. The release-date chart suggests six doublings in fitted frontier relative strength from π0 to GPT-6, averaging about four months per doubling through September 2026. This describes the index’s comparison scale, not a measured doubling of general physical ability every four months.
This is the first version of the index. Its methodology and per-benchmark coverage are public; Playground says the full 159,372-row dataset will be released later for reproduction. The charts show the announced snapshot, and the live leaderboard may change as new results and corrections arrive.
