GigaBrain-0.7 Tops All 4 RoboColiseum Benchmarks Days After Launch

机器之心 · wechat · 2026-08-21

GigaAI unveiled its new embodied foundation model GigaBrain-0.7 at WRC, sweeping all 4 sub-leaderboards on Zhiyuan's RoboColiseum simulation benchmark just five days after the platform launched, and beating leading open-source models on identical real-robot benchmarks. Code and weights will be open-sourced.
Introduces System-3: the GigaWorld-1 world model runs inside the robot's real-time decision loop, generating future action-value estimates and visual predictions before execution.
Three-layer algorithm pyramid: a VLM backbone (System-2) with temporal-spatial attention for long-horizon reasoning; MoT + FlowMatching action generation with shared parameters across embodiments; three-stage experience reinforcement (SFT → offline RL → human-in-the-loop online RL) approaching 100% success.
Five-layer data pyramid: 300M image-text pairs, 11,200 hours of sensor-identical human data, 5,000+ hours of synthetic/simulation data, 20,500 hours of real-robot data; only 300 UMI demonstrations teach MakerH01 a new task.
Experiments validate embodied scaling laws, with the pretrained base model handling many tasks at 100% success without per-task finetuning, plus a 20+ minute, 10-task one-take demo.

More from Embodied