Spirit AI beats Nvidia on RoboArena robotics benchmark

TL;DR

TNW news piece reporting that Hangzhou-based startup Spirit AI’s Spirit v1.6 foundation model for embodied intelligence scored 1,924 on the RoboArena leaderboard on 2026-06-03 (Wednesday), unseating NVIDIA’s Cosmos3-Nano-Policy at 1,881 — two days after Cosmos 3 was launched at Computex Taipei on June 1. NVIDIA’s DreamZero placed third at 1,763. The article frames this as the first time a Chinese model has topped RoboArena (a benchmark NVIDIA co-developed with Stanford and UC Berkeley) and contextualizes the result against a broader pattern of Chinese leadership across physical-AI benchmarks (WorldArena, WorldScore) and an extraordinary funding pace — Spirit AI closed a 1.5B yuan / $222M round the same day, its fourth in three months.

Key claims

Spirit AI’s Spirit v1.6 scored 1,924 on the RoboArena leaderboard, edging out NVIDIA’s Cosmos3-Nano-Policy at 1,881 and DreamZero at 1,763; Cosmos 3 had held the top spot for two days following its June 1 Computex launch.
RoboArena evaluates generalist robot policy on real-world action competence — object manipulation, navigation, tool usage, perception, planning, adaptability in unfamiliar environments — rather than chatbot fluency or image-generation quality.
The article distinguishes policy capabilities (act on observations — what RoboArena measures) from world capabilities (simulate and predict next-state given an action), and notes the field is converging via unified “Policy World Model” architectures.
Across adjacent benchmarks: WorldArena’s policy-evaluator track is led by Manifold AI’s WorldScape-0.2 (over NVIDIA Cosmos-Predict 2.5); WorldArena perception track is led by AgiBot’s GenieEnvisioner-Sim2.0-2B; WorldArena data-engine track by DexForce; WorldScore (text→world generation) led again by Manifold AI’s WorldScape-0.2 over WonderJourney.
Spirit AI raised 1.5B yuan (~ $222M) on the same day as the RoboArena announcement — its fourth round in three months — bringing valuation past 10B yuan ($ 1.4B); XYZ Embodied AI (BAAI-incubated) closed its pre-A having raised 1B yuan in 10 months; Manifold AI completed five rounds in 10 months.
Chinese robotics venture funding reached $3.4B in 2025, 42% more than the United States, with the gap reportedly widening in 2026.
NVIDIA’s counter-strategy at Computex: humanoid reference design partnership with Unitree (H2 Plus body, Sharpa Wave tactile hands, Jetson AGX Thor T5000) and the Cosmos Coalition enlisting Agile Robots, Black Forest Labs, Runway, and Skild AI to advance open world models.
Article cites Jensen Huang (“for robotic systems and physical AI, data is the hardest problem”) and an attributed-but-unverified Alexandr Wang claim that US companies rely on Chinese data to train robotics foundation models — framing China’s manufacturing supply chain + state-backed “data factories” as a structural data advantage.

Method

This is a tech-press news article, not a technical paper. The reporting is structured around (1) a benchmark scoreboard event (Spirit v1.6 > Cosmos3-Nano > DreamZero on RoboArena), (2) a survey of adjacent benchmark leaderboards (WorldArena policy/perception/data tracks, WorldScore), (3) a funding-pace recap for several Chinese embodied-AI startups, and (4) NVIDIA’s organizational response (Cosmos Coalition, Unitree partnership). No methodology or architectural details for Spirit v1.6, WorldScape-0.2, or GenieEnvisioner-Sim2.0-2B are disclosed in the piece — the linked sources are vendor blogs and a RoboArena arXiv page (arxiv.org/html/2506.18123v1).

Results

RoboArena leaderboard (at time of writing): 1. Spirit v1.6 (Spirit AI) — 1,924. 2. Cosmos3-Nano-Policy (NVIDIA) — 1,881. 3. DreamZero (NVIDIA) — 1,763.
WorldArena policy-evaluator track: WorldScape-0.2 (Manifold AI) leads, ahead of NVIDIA Cosmos-Predict 2.5.
WorldArena perception track: GenieEnvisioner-Sim2.0-2B (AgiBot) leads.
WorldArena data-engine track: DexForce leads.
WorldScore (text→world generation): WorldScape-0.2 (Manifold AI) leads, ahead of WonderJourney (Stanford + Google).
Funding: Spirit AI 1.5B yuan round on 2026-06-03, fourth in three months; >10B yuan valuation. XYZ Embodied AI: 1B yuan in 10 months. Chinese robotics 2025 VC: $3.4B (+42% vs US).

Why it’s interesting

Direct competitive datapoint against NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI, where Cosmos 3 was launched on June 1 with reported #1 on Artificial Analysis open-weights and “RoboArena” — within 48 hours a Chinese startup’s foundation model takes the top spot on that very benchmark, with no closed-flagship (e.g. Genie 3 / Project Genie (Project Genie: Experimenting with infinite, interactive worlds)) comparison shown in either direction. The piece is the wiki’s first market-signal entry pointing at the Chinese side of the open-WFM ecosystem (Spirit AI, Manifold AI, AgiBot, DexForce, XYZ Embodied AI), where prior filings have been NVIDIA-centric (NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry, Cosmos 3) or DeepMind-vertical (The Waymo World Model: A New Frontier For Autonomous Driving Simulation). Adjacent to RBench/RoVid-X (Rethinking Video Generation Model for the Embodied World (ReVidGen / RBench / RoVid-X)) on the robotics-benchmark dimension and to the consortium / data-factory pattern in EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World on the data-supply dimension — the article’s “data is the hardest problem” framing maps directly onto both.