Festivus

policy

PerAct

UW / NVIDIA · PyTorch

or hover any field below to flag it

Overview

Name
PerAct
Author
UW / NVIDIA
Framework
PyTorch
License
apache-2.0
Skill type
manipulation
Evidence level
verified
Task description
Perceiver-Actor for multi-task 6-DoF manipulation. Voxelizes workspace into 100x100x100 grid from multi-view RGB-D, processes with Perceiver IO transformer (2048 latents, 6 self-attention layers). Detects next-best voxel for keyframe actions. Strong inductive bias for 6-DoF from few demos per task.

Spaces

Action space
end-effector-pose · 8-dim · 1Hz
Observation space
  • type: multimodal
  • · front_rgbd (128x128)
  • · left_shoulder_rgbd (128x128)
  • · right_shoulder_rgbd (128x128)
  • · wrist_rgbd (128x128)
  • · language_instruction

Links

HuggingFace repo
null
Paper (arXiv)
https://arxiv.org/abs/2209.05451

Compatible robots

0+1 mentioned but not in catalog yet

No robots list PerAct as compatible yet. Know of one? Flag it above.

Compatible environments

2
tabletop-cleannot in seedtabletop-clutterednot in seed

Datasets that reference this policy

0

No datasets reference PerAct yet.