Daxi Robotics released ACE-Brain-0.6, an embodied foundation model with enhanced visual-spatial anchoring, fine-tuned from Qwen3.5-9B to improve object recognition, spatial localization, and interaction region judgment. Trained on about one million spatially anchored samples, it surpassed OpenAI GPT-6 Astra in six of eight Grounding benchmarks, scored 74.53 on seven public benchmarks, and 94.13 on the RobotGrounding real-robot evaluation.
On October 19, Daxiao Robotics released ACE-Brain-0.6, a visual-spatial-anchored embodied foundation model fine-tuned from Qwen3.5-9B to improve object recognition, spatial localization, and interaction-region judgment. Using about one million spatial-anchoring training samples, it outperformed OpenAI GPT-6 Astra on six of eight Grounding evaluations, averaged 74.53 across seven public benchmarks, and scored 94.13 on the self-built RobotGrounding benchmark.
On September 10th, DaXiao Robotics, in collaboration with Nanyang Technological University's S-Lab and other institutions, released and open-sourced the unified multimodal world model framework Puffin-World. This model treats physics, geometry, and appearance as three native 3D world states for camera pose understanding, free viewpoint generation, 3D reconstruction, and robot loop closure exploration. The team simultaneously released the code, model, and Puffin-16M dataset.