Embodied AI competition is shifting from model size to data depth. On August 25, the GigaAI team released training code, 3.5B Base weights and sample data for GigaBrain-0.7, all under Apache 2.0; the technical report landed on August 17, the GitHub repo has already collected 2.6k stars, and the paper topped Hugging Face Daily Papers with 90 upvotes the same day.
Three systems: a world model inside the decision loop
Vision-language-action (VLA) models are now the dominant paradigm for generalist embodied agents, completing complex long-horizon tasks in structured settings. Whether architecture still has headroom, how heterogeneous data can scale, and how far generalization reaches remain open questions. GigaBrain-0.7's answer is a three-system architecture that unifies understanding, prediction and action in one model. The most notable entry in the official highlights is System-3: it plugs a world model into the robot's real-time decision loop, letting it simulate and evaluate before acting — different from the common practice of using world models merely as offline data generators. Together with a dual-pyramid framework, a single pretrained model handles many task types out of the box: One Model, Many Tasks.
37.3k hours of heterogeneous data plus one-stage alignment
Pretraining scale is another anchor of this release: over 37,000 hours (37.3k per the README) of heterogeneous embodied data, spanning real-robot, UMI, egocentric, simulation, and world-model-generated sources. The alignment stage uses one-stage alignment training that jointly optimizes vision-language understanding and multi-embodiment action generation. Compared with the team's own GigaBrain-0 series and prior state-of-the-art models including π0.5, the authors report substantial gains in zero-shot capabilities, language-conditioned instruction following, and post-training task success rates; on the in-house Maker H01 platform and mainstream embodiments, coverage spans both home and industrial scenarios.
What gets open-sourced goes beyond weights
The GitHub repo lays out a complete post-training path: data in LeRobot format, norm-stats computation scripts with a delta-mask mechanism, eight built-in embodiment IDs (from AgileX and Agibot G1/dexterous hands to UMI, EgoDex and H01), and a training stack pinned to the PaliGemma2 environment (Python 3.11.10, with giga-datasets, giga-train and giga-models all locked at 1.1.0). On the deployment side, two profiles are provided for AgileX Cobot Magic and Maker H01, and the H01 client runs directly on Jetson. The repo even discloses loss numbers from a controlled ten-step smoke test: pick-and-place down 51.27%, push-buttons down 31.97%. Publishing smoke-test numbers in a README is not a common practice among embodied open-source projects.
Self-reported results and the to-be-verified list
The official highlights claim first place in all four evaluations on RoboColiseum (a benchmark with strong sim-to-real alignment), a wide lead in task success rates on Maker H01, and a single-take demo of over 20 minutes spanning more than 10 tasks, described as an industry first. A dose of skepticism: all these numbers are self-reported, and the benchmark code for RoboColiseum, RoboTwin2.0 and EBench, plus the VLM evaluation code, still sit in the TODO list — third-party replication has no entry point yet. For context, the team's GigaBrain-0.1 took first place on the RoboChallenge leaderboard in February 2026, and the team will host the GigaBrain Challenge at CVPR 2026 with three tracks: RoboTwin (simulation), GigaWorld (world model) and RoboChallenge (real robot).
So what
The rules of embodied foundation-model competition are now clear: compete on data hours, embodiment coverage, and world models in the loop. GigaBrain-0.7 does all three at once and open-sources everything — a ready starting point for teams wanting post-training on their own robots. But until the benchmark code lands, let the "first in all four" claims fly for a while before you buy them.
Source: arxiv.org/abs/2608.15875; code: github.com/open-gigaai/giga-brain-0