RoboColiseum, an embodied AI evaluation platform co-founded by Maniformer, Hugging Face, NVIDIA and Tsinghua University, said Claude Opus 5.5 ranked first among four flagship general-purpose models in its latest round of testing, with GPT-6 Astra second.
GPT-6 Astra led in instruction following and overall disturbance adaptation, scoring 99.35 on instruction following, according to the results released on Oct. 20. Opus performed better in spatial reasoning and general manipulation, and took the top spot among flagship models with the lowest total token consumption across all evaluation rounds. GPT-6.1 Sol lagged the other three models in image-quality perturbation tests, the platform said.
Despite Opus’s lead among general models, the top two positions in the overall ranking were still held by embodied policy models, which maintained an advantage in general manipulation.
The comparison offered a direct reference for the industry on where general-purpose large models and embodied policy models each excel and fall short, RoboColiseum said.
RoboColiseum evaluates models across four capability dimensions — instruction following, spatial reasoning, disturbance adaptation and general manipulation — through 78 high-fidelity simulation tasks. It records task completion steps and failure points to help developers locate capability gaps quickly. Maniformer, as a co-founder, supplied high-fidelity object assets and simulation scenarios for complex robot manipulation tasks.
The platform said its simulation results showed 89.5% correlation with real-robot deployment, with the sim-to-real gap for the same model at less than 10%, offering a reference for model screening and problem diagnosis before physical testing.
RoboColiseum, named after the Roman Colosseum, positions itself as an open, neutral and reproducible “arena” and “training ground” for embodied AI models. Since launching in August, more than 100 teams in China and abroad have used the platform for model training and evaluation, it said. The platform is now open to global developers.
Platform: https://robocoliseum.ai/
Technical report: https://robocoliseum.ai/techreport
AGIBOT
Figure AI
UBTECH