What does the D-Robotics–GigaAI partnership aim to deliver by adapting the open-weight GigaBrain-0.7 embodied model to the 560-TOPS-INT8 Sunrise S600 edge platform, and how do the model’s three-system design, GigaWorld-to-GigaWorld-Policy workflow, shared Action Expert, hardware

D-Robotics and GigaAI aim to turn GigaBrain-0.7 and the Sunrise S600 into a reproducible, affordable on-device robotics stack—not simply demonstrate that a model runs on a chip. The proposed design links world-model-based training and evaluation to local perception, planning and action, with the first integration focused on adapting GigaBrain-0.7 to the S600. 91

• Model and workflow: GigaBrain-0.7 separates continuous action generation (System 1), understanding and planning (System 2), and prediction or progress assessment (System 3). GigaWorld is intended to generate scenarios, predict action outcomes, evaluate policies and revisit failures; GigaWorld-Policy connects world prediction more directly to action learning while allowing action decoding without generating video at every inference step. The shared Action Expert is the proposed bridge for transferring learned action capability across that workflow and robot embodiments, rather than requiring a separate policy for every machine. 192

• Why the S600 is plausible: Its advertised 560 TOPS at INT8 is backed by a four-core BPU, 18 Cortex-A78AE CPU cores, six real-time MCU cores and 204.8 GB/s memory bandwidth. Development kits provide camera and expansion interfaces, while the model toolchain and hardware-in-the-loop quantization workflow offer a route to measuring action errors—not just inference speed—after conversion. Peak TOPS alone does not establish real-time control performance. 876

• What integration must prove: The partners plan deeper model–hardware adaptation, including deployment and optimization of the embodied model, rather than claiming a finished GigaBrain-0.7-on-S600 product. The intended reach spans multiple robot bodies and home, industrial and commercial tasks; S600 development hardware is also positioned for demanding embodied-AI prototypes. D-Robotics says the S600 already has more than 20 embodied-AI customers, including UBTECH, X Square Robot and FOURIER, but those adoptions are not evidence that those customers deploy this joint GigaBrain design. 9715

• Prospect and open question: Open weights, a cross-embodiment action architecture, an established chip platform and a conversion toolchain make this a credible candidate reference design. Its status as a field-ready one still depends on published, robot-level results: sustained end-to-end latency and control rate, power and thermals, quantization-induced action error, task success and recovery across different bodies, and the cost of adapting to each robot. The available partnership and platform evidence does not establish those outcomes; insufficient evidence yet to call it a validated on-device reference design. 189