Multiple trajectory subsets from the D4RL benchmark across OpenAI Gym environments like HalfCheetah, Hopper, and Walker2d. These files contain sequential transitions of states and actions specifically curated for training Decision Transformer models.
Use Cases
- Train a Decision Transformer model using the observations and actions sequences to predict next-step actions.
- Calculate returns-to-go for reward-conditioned policies using the rewards and terminals columns.
- Benchmark offline reinforcement learning algorithms against the actions recorded in the replay buffer.
Strengths
- Includes trajectory sequences containing observations, actions, and rewards columns.
- Derived from the D4RL benchmark covering standard OpenAI Gym continuous control tasks.
- Provides data formatted as a subset of the D4RL 'replay' category for offline RL training.