Open Locomotion Skills Dataset provides expert, random, and domain-randomized trajectories for training and evaluating locomotion policies. The dataset contains approximately 4,200 episodes and over 2.7 million steps, collected from MuJoCo/Gymnasium environments for four robot morphologies: HalfCheetah, Ant, Walker2d, and Hopper. It was created by author kanishqgandharv and last updated on March 27, —.
Use Cases
- Training offline RL locomotion policies based on expert and random trajectories
- Benchmarking policy performance across diverse robot morphologies mentioned in the description
- Studying the effects of domain randomization on locomotion skill transfer
Strengths
- Contains over 2.7 million steps of simulation data
- Includes trajectories for four distinct robot morphologies: HalfCheetah, Ant, Walker2d, Hopper
- Provides multiple quality tiers, including expert and random trajectories
Limitations
- Column-level documentation is absent; field semantics must be inferred after download
- Row count is unknown, which may limit suitability assessment
Provenance
- Source
- huggingface
- Collection Method
- Data collected from MuJoCo/Gymnasium environments.
- Time Range
- null
- Freshness
- Last updated —; freshness should be verified
- Geography
- null