Pusht Lewm Fr3: Real-Robot Planar Push Demonstrations with Visual Input

411 episodes of real-robot planar push demonstrations teleoperated on a Franka FR3 arm, totaling 93,728 frames recorded at 10 Hz. The dataset, created by Rongxuan-Zhou, contains raw pixel observations from a top-down D455 camera and auxiliary data from a wrist-mounted D405 camera. It was last updated on June 3, 2026.

Use Cases

• Train imitation learning policies based on visual observations from the D455 agent-view camera.
• Develop models for robotic state estimation using the recorded auxiliary camera data.
• Benchmark reinforcement learning algorithms on real-world planar push manipulation tasks.
• Study visual feature representation for robotic control from pixel inputs.

Strengths

• Contains 411 distinct demonstration episodes, providing multiple task executions.
• Includes 93,728 frames of visual data, captured at a consistent 10 Hz frequency.
• Uses exposure-locked cameras to ensure the white pushing rod remains visually distinct from the wooden background.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset contains no force/torque sensor data, which may limit certain physical modeling applications.

Provenance

Source

huggingface

Collection Method

Real-robot demonstrations teleoperated via WebXR/Quest on a Franka FR3 arm.

Freshness

Last updated 2026-06-03 02:29:03; freshness should be verified.