Core Concepts
FOURTRAN leverages SE(d) × SE(d) symmetry for efficient robotic manipulation in 3D environments.
Abstract
ABSTRACT:
Proposed Fourier Transporter (FOURTRAN) for pick-place tasks in robotics.
Achieves high sample efficiency using SE(d) symmetry.
Utilizes fiber space Fourier transformation for memory-efficient computation.
INTRODUCTION:
Imitation learning in SE(3) crucial for robotic manipulation.
Sample efficiency key due to complex tasks in 3D environments.
ACTION-CENTRIC MANIPULATION:
Vision-based policies struggle with deformable objects.
Action-centric manipulation efficient for dense output maps.
SYMMETRIES AND ROBOT LEARNING:
Translational equivariance improves learning efficiency.
Equivariant networks enhance sample efficiency.
METHOD:
PROBLEM STATEMENT:
Behavior cloning for pick-and-place tasks with expert demonstrations.
SE(d)-EQUIVARIANT PICK:
Pick network encodes pick pose distribution over SE(d).
SE(d) × SE(d)-EQUIVARIANT PLACE:
Place network infers place action conditioned on pick action.
SAMPLING ROTATIONS IN A COARSE-FINE FASHION:
Coarse-to-fine sampling method to improve memory efficiency and performance.
EXPERIMENTS:
MODEL ARCHITECTURE DETAILS:
Residual networks used with U-net backbone for fpick and fplace.
3D PICK-PLACE:
FOURTRAN outperforms baselines on RLbench tasks with high sample efficiency.
2D PICK-PLACE:
FOURTRAN achieves higher success rates than baselines on Ravens Benchmark tasks.
CONCLUSION:
FOURTRAN demonstrates significant improvements in sample efficiency and success rate.
Stats
Tests on RLbench benchmark achieve state-of-the-art results across various tasks.
FOURTRAN outperforms baselines by a margin of between six percent (STACK-WINE) and two-hundred percent (STACK-CUPS).