A robot VLA model that ships weights, code, and benchmarks

What it does

OpenDM is Dexmal’s open-source release of DM0.5, a Vision-Language-Action model built for open-world robot control across multiple embodiments. The repository includes the base pretrained weights, several task-specific fine-tuned checkpoints, and the full training and inference stack for researchers who want to adapt or deploy embodied agents. It targets long-horizon tasks, open-ended language instructions, and dynamic disturbances rather than narrow, scripted behaviors.

The interesting bit

Most VLA research stops at the architecture diagram; OpenDM publishes the weights, the LoRA workflows, and the evaluation guides for physical robots like the AgileX COBOT Magic. The README even includes head-to-head benchmark tables against Pi0, Pi0.5, and GROOT-N1.7 across simulated and real-world manipulation tasks, which is rarer than it should be in this field.

Key highlights

• Releases base DM0.5 weights plus fine-tuned variants for LIBERO, RoboTwin2.0, VLA-Arena, RoboDojo-Sim, and real-world Table30v2 tasks.
• Ships with both full supervised fine-tuning and LoRA workflows for adapting to new robots or tasks.
• Offers a fast inference backend built on TensorRT and Triton kernels, though it requires a specific PyTorch environment with FlexAttention support.
• Provides HTTP inference endpoints (/v1/infer and a legacy /process_frame path) for integration with external robot stacks.
• Includes hardware guides for physical platforms such as the AgileX COBOT Magic and DOS-W1, covering camera modifications and robot naming conventions.

Caveats

• The RoboDojo-Sim leaderboard scores come from the released DM05-MEM-Robodojo-Sim generalist checkpoint; the linked training guide explicitly notes its settings do not reproduce those table scores.
• Full supervised fine-tuning can exhaust memory on lower-end GPUs such as the RTX 4090 during FSDP initialization, so the project recommends LoRA fine-tuning for modest hardware.
• The fast inference backend is not a simple acceleration toggle: it requires TensorRT, Triton, and PyTorch FlexAttention, and the first launch for each checkpoint spends extra time building engines before the service is ready.

Verdict

Robotics researchers and embodied-AI teams who need a working VLA baseline with published weights and evaluation protocols should look here. If you are just looking for a drop-in robot controller without handling CUDA, FSDP, and TensorRT dependencies, this is still very much a research-grade stack.