Multitask DiT · REBOT Machine Tending II EE LBM (SieloRobotics)
A Multitask DiT policy from SieloRobotics for tending a 3D printer with a reBot B601 follower. It uses three RGB views and a 10-component robot state to generate relative end-effector actions for loading, unloading and sorting two parts.
The two dataset task variants swap the order of the triangular plastic part and the metal bracket. The printer door is opened and closed during loading and unloading, and the parts go to their corresponding bins. The source dataset contains 220 episodes and 525,493 frames at 30 fps. The training configuration excludes 26 episodes and sets a 5% evaluation split. LeRobot 0.7.1 uses flow matching with two observation steps, a 32-step prediction horizon, eight executed actions per chunk and 10 Euler integration steps. Cameras are top-left, top-right and wrist (640 × 480). Training is configured for BF16, batch size 192, Adam, a learning rate of 8.660254 × 10⁻⁵, and seed 1000.
Official resource (18)
Papers and technical reports
Paper / report unconfirmed
Official introductory materials are recorded; no version-specific paper or technical report has been verified.
Last checked: 2026-11-21
- Official source · huggingface.co ↗
Source publication date: 2026-11-21
- Sources checked · huggingface.co ↗
- Sources checked · huggingface.co ↗
- Sources checked · huggingface.co ↗
- Sources checked · huggingface.co ↗
- Sources checked · huggingface.co ↗
- Sources checked · www.sielorobotics.com ↗
- Sources checked · huggingface.co ↗
- Sources checked · arxiv.org ↗
- Sources checked · huggingface.co ↗
- Sources checked · huggingface.co ↗
How to access and use
- Model files
- Public access
- Training data
- Public access
Training code for this version is unconfirmed. Public code resources do not establish access to the exact training implementation.
Training / fine-tuning code
Model files
| File | Access |
|---|---|
| 036000 · Standard weights | Access |
| 036000 · EMA weights | Access |
| 036210 · Standard weights | Access |
| 036210 · EMA weights | Access |
SourceVerified on · 2026-11-22
Checks before deployment
Text conditioning is disabled. The publisher identifies the reBot B601 follower without specifying DM or RS and provides no physical-robot evaluation results. The dataset license is Apache-2.0. The model weights have no declared license. 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
Architecture and inputs / outputs
- Model type
- Action policy
- Architecture
- Diffusion policy · Transformer
- Weight parameters
- 511003530 parameters
- Action generation
- Flow matching
- Input
- RGB images / Robot state
- Action generation · Input modalities · Main category · Architecture: 036000 · policy configuration
- Action generation · Input modalities · Main category · Architecture: 036210 · policy configuration
- Architecture: Publisher model card
Scope and training
Ten action components: three position coordinates, six rotation-representation values and one gripper command.
horizon 32 · quantiles_clipped
Embodiment scope
Single embodiment
Functional purpose
action_control
Output
Actions
Robot form
Robotic arm
Training
Imitation learning
- Embodiment scope · Robot form · Training: Publisher model card
- Embodiment scope · Robot form · Ten action components: three position coordinates, six rotation-representation values and one gripper command. · continuous: Dataset robot and feature metadata
- Functional purpose · Output · Ten action components: three position coordinates, six rotation-representation values and one gripper command. · continuous: 036000 · policy configuration
- Functional purpose · Output · Ten action components: three position coordinates, six rotation-representation values and one gripper command. · continuous: 036210 · policy configuration
- Robot form: LeRobot reBot B601 hardware guide
- Training: 036000 · training configuration
- Training: 036210 · training configuration
- Ten action components: three position coordinates, six rotation-representation values and one gripper command. · continuous: LBM DiT implementation patch