ME297SJSU2026 ACT SO-101 Dial Pick-and-Place Mix v3 DAgger2

Third-party fine-tune
Papers and technical reports

The ME297SJSU2026 course team released this LeRobot ACT fine-tune for SO-101 dial pick-and-place in MuJoCo. Checkpoint 010000 was trained on SJSU College of Engineering HPC using a configuration that lists 1,220 episodes. It takes three camera images and a six-dimensional state and returns chunks of six-dimensional actions.

View full model description

Teaching and research on simulated SO-101 dial pick-and-place.

Papers and technical reports

Paper / report unconfirmed

Official introductory materials are recorded; no version-specific paper or technical report has been verified.

Last checked: 2026-11-18

How to access and use

Model files
Public access

Training code for this version is unconfirmed. Public code resources do not establish access to the exact training implementation. View verification notes

Training / fine-tuning code

Model files

Model files
ME297SJSU2026 ACT SO-101 Dial Pick-and-Place Mix v3 DAgger2 · 010000Access
Release information

SourceVerified on · 2026-11-18

Checks before deployment

The v3 real-rig profile describes the simulator setup. Evidence of physical-robot execution or a success rate for this checkpoint was unavailable. License, commercial permissions, dataset composition, and code access remain unverified. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Deployment and usage resources

Runtime and frameworks
LeRobot / PyTorch
Fine-tuning requirements
20,000 configured steps; batch size 32

Architecture and inputs / outputs

Model type
Action policy
Architecture
Transformer
Backbone
ResNet18
Weight parameters
51668614 parameters
Input
RGB images / Robot state
Output
6-dimensional actions
Released
November 19, 2026

Scope and training

Base model

LeRobot ACT

Embodiment scope

Single embodiment

Functional purpose

Action control

Record identity

Third-party fine-tune

Output

Actions

Pipeline task

Robotics

Prediction output

100 actions per chunk

Robot form

Robotic arm

Task scope

Single task

Training data scale

1,220 configured episodes

Training

Imitation learning

Related robots

Full parameters and verification notes (model card)

The ME297SJSU2026 course team released this LeRobot ACT fine-tune for SO-101 dial pick-and-place in MuJoCo. Checkpoint 010000 was trained on SJSU College of Engineering HPC using a configuration that lists 1,220 episodes. It takes three camera images and a six-dimensional state and returns chunks of six-dimensional actions.

The DAgger2 run initialized from the mix v3 DAgger1 checkpoint at step 15,000. Both DAgger2 and DAgger3 configure 20,000 steps, batch size 32, seed 1000 and learning rate 0.00001. DAgger3 initializes from this DAgger2 checkpoint, lists 1,528 episodes, and sets eval_steps to 1,000 versus 0. The DAgger2 episode list alone does not establish the mix of original and correction data. The main SafeTensors file is 206,699,768 bytes and Hugging Face reports 51,668,614 F32 parameters. The repository includes policy processors and their normalization statistics.

The training configs identify the local datasets as local/so101_pick_place_rig_v3_dagger2 for DAgger2 and local/so101_pick_place_rig_v3_dagger3 for DAgger3. The two policy configs differ only in their initialization path. The configured evaluation split is 0.04877 for DAgger2 and 0.03894 for DAgger3. The configurations do not establish shared trajectories or public dataset access. Both main weight files are 206,699,768 bytes and have different SHA-256 hashes. On 2026-11-19 Taipei time, the repository was created at 02:08:33, the weights were uploaded at 02:08:41, and the model card was added at 02:08:42. SHA-256 is 6df223d9b41106cd033d0e9db67a980c3300b49d2fd1574b9c3a387106f73d92.

Submit a source or correction