pusht_v2_keypoints (TsFile)

Apache TsFile version of cadene/pusht_v2_keypoints.

Overview

The keypoints variant of the PushT benchmark, formatted with LeRobot. A 2D pointer agent pushes a T-shaped block onto a T-shaped target region. Each frame records the agent state, the action, the environment keypoint state, and the step reward / success flag.

• Robot: 2d pointer (2D agent).
• Task: "Push the T-shaped blue block onto the T-shaped green target surface." (1 task).
• Scale: 206 episodes, 25,650 frames, single train split.
• Sampling rate: 10 fps.
• Videos: none — this dataset stores keypoints, not images.

Schema (TsFile structure)

• Time (INT64, milliseconds) — round(timestamp * 1000), restarts per episode.
• episode_index (TAG) — episode device dimension; query one episode with WHERE episode_index=0.
• task_index (TAG) — task device dimension (single task here).
• frame_index (FIELD, INT64) — original per-episode frame counter.
• sample_index (FIELD, INT64) — source index column, renamed.
• next_reward (FIELD, FLOAT) — step reward (source next.reward).
• next_success (FIELD, BOOLEAN) — step success flag (source next.success).
• observation_state_0..1 (FIELD, FLOAT) — 2D agent position.
• action_0..1 (FIELD, FLOAT) — 2D action.
• observation_environment_state_0..15 (FIELD, FLOAT) — the 8 (x, y) keypoints describing the T block / agent geometry.

Vector columns are flattened by preserving the source name (. → _) and appending the element index. The source timestamp column is dropped because it equals Time ÷ 1000 seconds.

Usage

Read the .tsfile files with the Apache TsFile Java or Python SDK.

Source & license

• Original dataset: https://huggingface.co/datasets/cadene/pusht_v2_keypoints
• Author / publisher: cadene (LeRobot / Hugging Face)
• Benchmark: PushT — https://diffusion-policy.cs.columbia.edu/
• Created with: LeRobot
• License: apache-2.0