Open Source RoboticsRobots, software & learning

ACT: Action Chunking with Transformers

Software

Action-chunk imitation-learning implementation with paired-arm simulation tasks and separate ALOHA hardware integration. [1] [2] [3] [4] [5] [6]

Installation & first result →Official websiteGetting started

For
Robot-learning developers investigating imitation policies [1] [2] [3] [4] [5] [6]
Source rights
Named licenses are documented by part. Review licenses
Unknowns
operating system not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

ACT implements action-chunk imitation learning with Transfer Cube and Bimanual Insertion simulation tasks. Its real-robot route additionally requires the separate ALOHA project. This entry concerns policy and simulation code; it does not establish another open hardware design or a reproduced manipulation result. [1] [3]

Is this for you?

Known limitations
  • Dataset/checkpoint normalization and14-dimensional paired-arm state must match.
  • Real-robot route requires separate ALOHA installation; no hardware operation or success rate reproduced.

Editorial assessment of cited source; no runtime or hardware test.

[1] [2] [3] [4] [5] [6] · Editorial assessment

What you need

Dependencies
  • README Python3.8.10,MuJoCo2.3.7,dm_control1.0.14 and editable act/detr install
  • PyTorch/torchvision and other packages mostly unpinned; inspected training/evaluation uses CUDA

[1] [2] [3] [4] [5] [6]

Compute for the workload
  • Selected training/evaluation source invokes CUDA; no independently validated CPU-only path or measured minimum GPU/RAM budget.

Upstream documented requirement/configuration only, not independently tested installation or compatibility.

[4]

Installation & first result

Open the upstream getting-started guide [1] [2] [3] [4] [5] [6]

  • Inspect sim_transfer_cube_scripted and replace DATA_DIR in constants.py.
  • Follow the pinned environment/data-recording route, then keep dataset_stats.pkl with the trained checkpoint.

Editorial assessment of cited source; no runtime or hardware test.

[1] [2] [3] [4] [5] [6] · Editorial assessment

Installation

Conda Python 3.8.10; README pins mujoco 2.3.7 and dm_control 1.0.14, installs PyTorch/torchvision and other dependencies then editable act/detr. Match CUDA environment before training/evaluation.

Upstream documented requirement/configuration only, not independently tested installation or compatibility.

[1] [4]

Project notes

Start with the documented simulation

The README creates a Python 3.8.10 conda environment, pins mujoco 2.3.7 and dm_control 1.0.14, installs PyTorch/torchvision and other dependencies, then installs act/detr in editable mode. Most of those other packages are unpinned. The inspected training/evaluation source calls CUDA, so the listed environment commands are not evidence of a CPU-only route or a complete compatible GPU stack. No environment was installed here. [1] [4]

A bounded starting task is sim_transfer_cube_scripted. The README records 50 scripted episodes with record_sim_episodes.py and a selected dataset directory before training through imitate_episodes.py. constants.py still contains a DATA_DIR placeholder; its task configuration names 50 episodes, a 400-step episode and the top camera. Replace paths and inspect the task before attempting the commands. These are source instructions, not data collected by this review. [1] [5]

A checkpoint needs its data statistics

The selected training source fixes state_dim to 14, configures the query count from chunk_size and saves dataset_stats.pkl beside the checkpoint. During evaluation it restores those statistics and converts predicted normalized actions using action_std and action_mean. utils.py separately normalizes joint positions and actions from the training data. Retain that exact task/action convention and statistics when moving a checkpoint; exchanging only a weight file does not establish meaningful commands for another embodiment. [4] [6]

The documented example uses chunk_size 100 and exposes temporal aggregation as an additional flag. Those choices belong to a specific policy/configuration, not an assurance of smooth physical motion. The README's example success rates and training tips remain author reports; no inference, rollout, normalization check or benchmark was run here. [1]

The repository's MIT notice concerns its code. Linked demonstrations, simulator assets and ALOHA hardware have separate sources and rights. No external demonstration payload or pretrained checkpoint was validated, and no real robot interface was connected. [1] [2]

What we checked

Source review only; no environment, dataset, policy inference/training or hardware connection executed.

Not yet established

  • Requirements: operating system, host computer
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Not established
Latest release
Not established
Profile updated
Nov 1, 2026
Sources last checked
Nov 1, 2026
Next review
Dec 1, 2026
Reviewed revision
742c753

Source files and licenses

Distinct imitation-policy implementation and simulation workflow; separate from the ALOHA mechanical platform and external demonstrations.

LayerAvailability / license
softwareavailable · MIT
MIT code notice; external demonstrations, robot assets and hardware sources require their own rights review. [1] [2] [3] [4] [5] [6]
documentationavailable · MIT
MIT code notice; external demonstrations, robot assets and hardware sources require their own rights review. [1] [2] [3] [4] [5] [6]

No separate physical design, pretrained distribution or dataset is selected as the released contribution in this scoped software review.

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history