機器人與具身 AI
GigaBrain-0.7 Releases 3.5B VLA Weights and Training Code, Covering Eight Robot Configurations
GigaAI has released GigaBrain-0.7’s base weights, training pipeline, and robot deployment examples. The model was pretrained on 37,300 hours of heterogeneous embodied data. However, the paper’s three-system architecture and leading results are not yet fully reproducible, and the official evaluation code remains on the roadmap.
On August 25, GigaAI added the GigaBrain-0.7 code, 3.5B base model, and sample data to its existing open-source project, providing the first downloadable and trainable implementation of the technical report published on August 16. The released model uses PaliGemma2 as its vision-language backbone, paired with a continuous-action expert interleaved layer by layer. Official ablation results show that this dual-stream integration outperformed variants that read only the final layer or used layer-wise cross-attention across three real-robot tasks, although inference time increased from 0.073 seconds for the fastest variant to 0.221 seconds.
The complete design divides capabilities among three systems: System 1 generates low-level actions; System 2 uses historical images to track progress and decompose long-horizon tasks; and System 3 uses a world model to predict future images and estimate state values. The pretraining dataset totals approximately 37,300 hours and includes trajectories from real robots, UMI, egocentric data, simulation, and world-model generation, spanning 16 robot embodiments and roughly 270 million vision-language samples. The released pipeline uses the LeRobot format and handles different control spaces through embodiment IDs, action masks, and per-dataset normalization statistics. The repository already provides training, inference-server, and ROS-client examples for AgileX Cobot Magic and Maker H01.
In the paper, GigaBrain-0.7 reports RoboColiseum scores of 0.8166 for instruction following, 0.4729 for spatial reasoning, 0.6800 for robustness, and 0.6092 for general manipulation, exceeding all listed public baselines in each category. However, these comparisons were compiled by the authors, and some real-robot results involved additional fine-tuning on task-specific demonstrations, so they should not be interpreted as fully zero-shot capabilities. More importantly, the repository still lists the VLM, RoboColiseum, RoboTwin 2.0, and EBench evaluation code as forthcoming. The current release also includes only sample data rather than the complete 37,300-hour corpus. The engineering community’s next priorities should be to determine whether the three-system inference architecture can be reconstructed end to end from the released components and whether third parties can reproduce the reported results using the same robots and benchmark protocols.