EmbodiedWorld-200K
A 200K-scale multi-level annotated dataset for open-world embodied agents. Each sample contains low-level annotations such as camera motion trajectories and high-level annotations such as semantic instructions, and a unified evaluation split is provided for offline pretraining and fair evaluation.
Timeline (last 90 days)
The research team released the EmbodiedWorld-200K dataset, containing more than 200,000 samples and covering multiple scenarios.
Each sample in EmbodiedWorld-200K comes with multi-level annotations from low-level to high-level, including low-level camera motion trajectories and high-level semantic instructions.
EmbodiedWorld-200K provides a unified evaluation split, including standardized data splits and task-specific metrics, for fair comparison across different methods.
Pretraining on EmbodiedWorld-200K significantly improves EWA's open-world embodied planning capability.
The team has publicly released the dataset, baseline model checkpoints, and training and evaluation code, with the project homepage at xiaokunfeng.github.io/EmbodiedWorld-200K.
All Facts (5)
Each sample in EmbodiedWorld-200K comes with multi-level annotations from low-level to high-level, including low-level camera motion trajectories and high-level semantic instructions.
50%UnverifiedEmbodiedWorld-200K provides a unified evaluation split, including standardized data splits and task-specific metrics, for fair comparison across different methods.
50%UnverifiedPretraining on EmbodiedWorld-200K significantly improves EWA's open-world embodied planning capability.
50%UnverifiedThe research team released the EmbodiedWorld-200K dataset, containing more than 200,000 samples and covering multiple scenarios.
50%UnverifiedThe team has publicly released the dataset, baseline model checkpoints, and training and evaluation code, with the project homepage at xiaokunfeng.github.io/EmbodiedWorld-200K.
50%