Aligning Policy Encoders with a Shape Teacher
Aligning Policy Encoders with a Shape Teacher
This research introduces shape-teacher alignment, a method that uses pretrained 3D shape representations as targets for training an image-based policy encoder. By leveraging Point-MAE features from individual objects to align with corresponding image patches, the approach transfers geometric structure to the policy without altering its deployment inputs or architecture. Experiments on the LIBERO benchmark show that this method produces more structured image representations around object shapes and parts, accelerates policy convergence by 4.7x on the Object suite and 5.6x on the Spatial suite, and improves success rates—all while maintaining a purely image-based interface at deployment without requiring depth input or the shape teacher during operation.
The paper proposes a training-time mechanism called shape-teacher alignment that transfers geometric structure from pretrained 3D Point-MAE features to an image-based policy encoder, enabling faster convergence and improved success rates on the LIBERO benchmark while maintaining a purely image-based deployment interface without requiring depth sensors or 3D inputs at runtime.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
t inputs or architecture. On LIBERO, shape-teacher alignment produces image representations that are more structured around object shape and corresponding parts, while substantially accelerating policy convergence and improving success rates. For pretrained policies, convergence is accelerated by 4.7x and 5.6x on the Object and Spatial suites, respectively, without requiring the shape teacher or depth input at deploy
Open source S4
ining while retaining a purely image-based interface at deployment. We propose shape-teacher alignment, a training-time mechanism that uses a pretrained 3D shape representation as an alignment target for an image-based policy encoder. Specifically, we use Point-MAE features from individual objects as targets for corresponding image patches, transferring geometric structure to the policy without changing its deploymen
Open source S3
Aligning Policy Encoders with a Shape Teacher
Open source S1
Source：arXiv Robotics — research abstracts · arxiv.org