Pengyue-polaron provides precomputed Pi3X geometry targets for the LIBERO robotics dataset. The dataset contains 1,768 episodes with 286,537 source frames from agent and wrist cameras, processed to a resolution of 224×224 and stored across 3,536 NPZ files totaling 230.0 GB. It was last updated on July 19, 2026.
Use Cases
Training robotic manipulation models based on precomputed geometric scene representations.
Benchmarking computer vision algorithms for 3D geometry prediction from multi-view images.
Developing sim-to-real transfer methods using processed real-world camera streams.
Strengths
Contains 1,768 distinct episodes, providing a substantial number of scenarios.
Includes data from two camera perspectives (agent and wrist), offering multi-view geometry.
Processed from 286,537 source frames, indicating a large volume of underlying visual data.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
The dataset size of 230.0 GB may present storage and download challenges.
Provenance
Source
huggingface
Collection Method
Source images and camera intrinsics from LIBERO were resized and processed with yyfz233/Pi3X.
Freshness
Last updated 2026-07-19 03:45:14; freshness should be verified.
License is unknown; users should verify permissions before use.