Kinetix-Offline: 3 Billion Expert Transitions for Physics Tasks

3 billion expert transitions across 11 million unique physics-based tasks collected in the Kinetix environment. The dataset contains offline expert trajectories from a JAX-based 2D rigid-body physics environment where tasks are procedurally generated. It was authored by mbeukman and last updated on Hugging Face in May 2026.

Use Cases

• Training offline RL agents based on the 3 billion expert transitions.
• Benchmarking imitation learning algorithms based on the expert trajectories.
• Studying generalization in procedurally generated environments based on the 11 million unique tasks.
• Developing physics-based control policies based on the described motor and thruster action space.

Strengths

• Contains 3 billion expert transitions, indicating a large-scale collection.
• Covers 11 million unique procedurally generated tasks, suggesting high task diversity.
• Trajectories are from specialist PPO agents, indicating a source of high-quality demonstrations.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Collected from specialist PPO agents operating in the Kinetix physics environment.

Freshness

Last updated 2026-05-19 08:27:50; freshness should be verified.