Sign in to view source links and access this dataset
Description
DexWild is a large-scale, high-quality dataset of human dexterous hand motions collected for robotics research. It contains 9,505 episodes across 5 tasks and 93 environments, totaling over 33 hours of data. The dataset was created by Tony Tao, Mohan Kumar Srirama, Jason Jingzhou Liu, Kenneth Shaw, and Deepak Pathak for the Robotics: Science and Systems (RSS) 2025 conference.
Use Cases
Training imitation learning policies based on high-quality human demonstrations.
Benchmarking dexterous manipulation algorithms across diverse tasks and environments.
Developing sim-to-real transfer methods using real-world human motion data.
Studying human hand motion primitives and strategies for complex object interactions.
Strengths
Large scale with 9,505 recorded episodes.
High diversity across 5 distinct tasks and 93 environments.
Substantial duration with over 33 hours of collected data.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Data may reflect task and environmental bias inherent to the collection setup.
Provenance
Source
Tony Tao, Mohan Kumar Srirama, Jason Jingzhou Liu, Kenneth Shaw, Deepak Pathak
Collection Method
Human demonstration data collection, as referenced in the associated project page and data collection guide.
Freshness
Last updated 2025-11-07 21:09:44; freshness should be verified.
License is unknown and must be verified before use.