Run inference with a pretrained DM0.5 checkpoint to predict robot actions from images and text instructions.
Fine-tune DM0.5 on a custom robot dataset by registering it with the project's dataset config format.
Evaluate DM0.5 against benchmarks like LIBERO and RoboTwin 2.0 using the included evaluation workflows.
| dexmal/opendm | marcj/papernews | python/tzdata | |
|---|---|---|---|
| Stars | 107 | 107 | 107 |
| Language | Python | Python | Python |
| Last pushed | — | — | 2026-07-01 |
| Maintenance | — | — | Active |
| Setup difficulty | hard | moderate | easy |
| Complexity | 5/5 | 2/5 | 1/5 |
| Audience | researcher | general | developer |
Figures from each repo's GitHub metadata at analysis time.
Requires an NVIDIA GPU, Docker with the NVIDIA container toolkit or a matching CUDA and flash-attn setup, and downloading large pretrained checkpoints.
OpenDM is the code and model release behind DM0.5, a Vision Language Action model built by Dexmal for controlling robots in open ended, real world settings. A Vision Language Action model, or VLA, takes in images and a text instruction and outputs the physical actions a robot should take to complete that instruction. DM0.5 builds on an earlier version called DM0, adding improvements aimed at handling open ended instructions, tasks that unfold over a long sequence of steps, unexpected disturbances during a task, and control of more than one type of robot body. The repository provides the model weights, training and inference scripts, examples for registering new datasets, and evaluation workflows so researchers and developers can train, fine tune, evaluate, and deploy the model themselves. Pretrained checkpoints are hosted on Hugging Face and ModelScope, including a base model meant for fine tuning and versions already fine tuned for two robotics benchmarks, LIBERO and RoboTwin 2.0. On those benchmarks the README reports strong average scores, close to ninety nine percent on LIBERO and around ninety three and a half percent on RoboTwin 2.0. Getting started requires a Linux machine running Ubuntu, an NVIDIA GPU, and either Docker with the NVIDIA container toolkit or a local Conda environment with PyTorch and flash attention installed. The project recommends the Docker route to avoid mismatched versions of CUDA and other dependencies. Once installed, a launcher script starts an inference server that loads a chosen checkpoint and exposes an HTTP endpoint. Client applications send it an instruction in text, the robot's current state, and one or more camera images, and the server responds with a sequence of predicted actions. For training, users register their own robot dataset by creating a small configuration file that describes where the data lives, which images to use, and how the robot's state is structured, following a template already included in the repository. The project is released under the Apache 2.0 License.
A Vision Language Action model and toolkit for training, fine-tuning, and deploying robots that follow text instructions from images.
Mainly Python. The stack also includes Python, PyTorch, CUDA.
Free to use, modify, and distribute, including commercially, as long as you keep copyright and license notices.
Setup difficulty is rated hard, with roughly 1h+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
double-check against the repo, no cap.