EgoNet
The largest most diverse egocentric dataset of human actions
About
EgoNet is the largest most diverse egocentric dataset of human actions. Building on previous data collection efforts like Ego4D, EgoNet expands upon both the diversity and size of the dataset while also adding necessary features including stereo RGBD data, hand pose tracking, and SLAM.
The name EgoNet, is a tribute towards the inspiration of ImageNet and Prof. Li Fei-Fei, Prof. Jia Deng, Prof. Olga Russakovsky, Prof. Alex Berg, and Prof. Kai Li.
Why EgoNet?
The EgoNet project was driven by two foundational needs in embodied AI and robotics research.
1.
First, the field of learning from humans lacks a clear North Star. While research spans everything from reinforcement learning to imitation learning, there's no canonical benchmark or dataset that unifies the community's efforts to bridge the embodiment gap and thereby leverage the incredible diversity and collection capability of egocentric human data.
2.
Second, there is a data bottleneck in the overall field of robotics that prevents at-scale deployment of robotics in real world settings. Foundation models are increasingly data-hungry, yet they are starved of the rich, diverse, real-world footage needed to generalize across tasks and environments. Simulations ignore many nuances of the physical world. Teleoperation is far too costly to reach internet scale even for the well capitalized enterprise labs. Human data presents a viable solution at capturing such diversity and scale.
EgoNet is designed to address both these gaps. It provides a shared challenge and a massive, flexible foundation for the next generation of robotics research. Much like how ImageNet sparked the deep learning revolution by making neural networks usable at scale, we believe EgoNet can do the same for embodied intelligence by inspiring research that ultimately bridges the embodiment gap.
Contributors From
