ALOHA Sim (Transfer Cube / Insertion)

The two bimanual simulated tasks — cube transfer and peg insertion — built in MuJoCo for the ACT paper.

The ALOHA simulated tasks are two bimanual fine-manipulation tasks built in MuJoCo by Tony Zhao and colleagues for the 2023 ACT (Action Chunking Transformer) paper, meant to let others reproduce the results easily. Transfer Cube requires the right arm to pick up a red block on the table and hand it off into the left arm's gripper, with roughly a 1 cm clearance; Insertion requires the two arms to each pick up a socket and a peg and mate them in mid-air, with roughly a 5 mm clearance. Each task ships with 50 scripted demonstrations and 50 human-teleoperated ones. Hugging Face packaged it as the gym-aloha environment (a 14-dimensional action space: 6 joints plus 1 gripper per arm), and provides the dataset through LeRobot, making it a common starting benchmark for learning imitation learning.

ExampleThe ACT project's own documentation notes that, trained on 50 scripted demonstrations, success rate should land around 90% for cube transfer and around 50% for insertion.

Also called

gym-aloha, Transfer Cube, Insertion, AlohaTransferCube-v0, AlohaInsertion-v0

Related

Action Chunking with Transformers · ALOHA · Bimanual Manipulation · LeRobot · MuJoCo (Multi-Joint dynamics with Contact) · Imitation Learning

Sources

Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT)
huggingface/gym-aloha (GitHub)
tonyzhaozh/act (GitHub)

As of

2026-09