Blogs›Skild AI›Omni-bodied Learning from Video
Skild AI's Omni-bodied Learning enables robots to acquire new skills by watching human videos, bridging the embodiment gap and overcoming the data bottleneck in robotics. This approach leverages abundant human video data and requires minimal robot-specific data for fine-tuning. This post details the development of an omni-bodied brain trained on a universe of 100,000 different robots, enabling end-to-end locomotion driven entirely by online vision and proprioception. The model demonstrates in-context learning, adapting to novel scenarios such as limb loss, broken legs, jammed wheels, and walking on stilts, showcasing resilience and zero-shot control capabilities.
1/11/2026
Introduces Omni-bodied Learning, a novel approach for robotics that allows models to learn new skills directly from human video demonstrations. Addresses the challenges of missing signals and the embodiment gap by mapping human actions to robot actuations, significantly reducing the need for robot-native data.
9/23/2025
Introduces the concept of an 'omni-bodied brain' trained on a vast simulated robot multiverse to achieve generalizable locomotion. Demonstrates in-context learning in robotics through various failure and adaptation scenarios (precarious scenarios, loss of limbs, broken legs, jammed wheels, walking on stilts), highlighting the model's ability to adapt to unseen morphological changes and environmental conditions without explicit fine-tuning.
8/5/2025
Introduces the low-level control capabilities of Skild Brain, enabling end-to-end locomotion driven entirely by online vision and proprioception. The single neural network directly outputs low-level motor commands from raw images and joint feedback, allowing robots to adapt dynamically to new terrain, climb stairs, and step over obstacles without prior planning or mapping. Demonstrates robustness, adaptability, and precise footwork in real-world scenarios, including carrying payloads.