Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

Dexmal DM0.5

Dexmal's open VLA robot model with 60 s of visual memory, in LeRobot

License

  • code (Dexmal/OpenDM): Apache 2.0
  • LeRobot dm05 policy code: Apache 2.0

Code: Apache-2.0, read from the LICENSE file of Dexmal/OpenDM. Weights: the Gemma licence on every DM0.5 model card in Dexmal's collection and on LeRobot's converted copies, none of them gated; the backbone is built on Google's Gemma 3. The Gemma terms allow commercial use, changes and redistribution, but anyone using or sharing the weights must follow Google's prohibited-use policy and pass these terms on. LeRobot: the dm05 policy code in LeRobot is Apache-2.0, and its guide says the weights follow their model cards. Datasets: Dexmal's RoboTwin 2.0, RoboDojo-Sim and SO-101 training sets on Hugging Face carry only a generic Creative Commons tag with no version or terms; check with Dexmal before reuse. Hosted service: Dexmal MaaS is a separate paid service under its own terms. Card image: a frame of Dexmal's own demo video, not covered by either licence. Checked 2026-11-12 against the LICENSE file, the Hugging Face model and dataset cards and the LeRobot guide.

Overview

DM0.5 is an open vision-language-action (VLA) model from the Chinese robotics company Dexmal, released on 9 July 2026 as the successor to its DM0 model of February 2026. It takes camera images, the robot's joint state and a natural-language instruction, and outputs chunks of continuous robot actions. The model pairs a 4B vision-language backbone built on Google's Gemma 3 with a 680M action expert trained with flow matching, and it can look back over key frames from up to the last 60 seconds, so it can remember where an object was or what a person demonstrated earlier in the task. Dexmal trained it on real manipulation data from several robot types, navigation data, first-person human video and general vision-language data. The OpenDM repository releases the weights, training and inference code, fine-tuning recipes and guides for real robots. Released checkpoints include the base model, a memory version, fine-tuned models for LIBERO, RoboTwin 2.0, VLA-Arena, RoboDojo-Sim, RoboChallenge Table30 v2 and an SO-101 pick-cube task. The code is Apache-2.0; the weights are under the Gemma licence. LeRobot added DM0.5 as the dm05 policy in October 2026.

DM0.5 is Dexmal's robot foundation model released on 9 July 2026, the successor to its DM0 of February 2026. The OpenDM repository holds the code; the weights are in a Hugging Face collection and on ModelScope. Dexmal describes the model in a technical blog rather than a paper.

DM0.5 is a VLA: a vision-language model that reads the camera images and the instruction, joined to an action expert that turns its features into motor commands. The backbone is a 4B model built on Gemma 3, and the action expert has 680M parameters. Each call produces a chunk of future actions (action chunking); by default 10 flow-matching steps generate 50 actions, and Dexmal reports about 10 calls per second on one RTX 4090 and 20 on one H100.

Dexmal names three additions over DM0:

- Memory: besides the current frame, the model sees compressed key frames from up to the last 60 seconds, so it can put a cup back where it was or follow a rule a person showed at the start. - Reasoning during training: 11 extra text tasks make the model describe task stages, upcoming events and its next actions, not only predict motor commands. - Trajectory alignment: predicted actions are matched to the demonstration by progress rather than by timestamp, so teleoperation data recorded at different speeds teaches the same motion.

Training data mixes real manipulation from AgileX ALOHA, Galaxea R1 Lite, AgiBot G1, Franka Emika Panda, UR5, ARX5 and Dexmal's own dual-arm mobile robot with navigation data, first-person video of people handling objects and general image and video data. Dexmal says the model can be post-trained for robot bodies it has not seen (cross-embodiment); the base checkpoint uses a 14-value state and action layout for two 6-joint arms with grippers, and the core model handles up to 32 values.

Released checkpoints:

- DM05, the base model for fine-tuning, and DM05-MEM, a version with 32 history slots. - Benchmark models for LIBERO, VLA-Arena and RoboTwin 2.0, the RoboDojo-Sim model for ARX X5 arms, and a RoboChallenge Table30 v2 collection for four real robot types. - DM05-SO101-Pick-Cube, a LoRA fine-tune for the SO-101 arm with its own training guide.

The repository has an inference server with an HTTP API, a fine-tuning guide for new robot data with full training or LoRA, benchmark recipes such as the RoboTwin 2.0 guide, and a guide to the real robots Dexmal used: the AgileX COBOT Magic and Dexmal's own DOS-W1, with their camera positions. Since LeRobot 0.7, DM0.5 is also the dm05 policy: the LeRobot guide trains from lerobot/dm05_base on any standard LeRobot dataset.

Dexmal reports 99.0% on LIBERO against 96.9% for pi0.5, scores of 93.6% and 93.3% on RoboTwin 2.0, and a 43% success rate on the real-robot RoboChallenge Table30 v2 against 14.3% for pi0.5. Dexmal also says DM0.5 ranked first on the RoboColiseum and RoboDojo-Sim leaderboards; these are Dexmal's numbers and claims. Dexmal also offers DM0.5 as a paid hosted service, Dexmal MaaS.

Why it matters

DM0.5 joins the open VLA models we list, such as pi0 and pi0.5, NVIDIA Isaac GR00T, LingBot-VLA, WALL-OSS and GalaxeaVLA. What Dexmal puts first is memory: DM0.5 is trained to use up to a minute of past views as well as the current frame, which matters for tasks where something important has gone out of sight. It also ships fine-tuned checkpoints for several benchmarks and four real robot types, plus an SO-101 example that owners of that low-cost arm can reproduce. Being a LeRobot policy means anyone with a LeRobot dataset can fine-tune it with the same tools as other policies, and Dexmal's own benchmark tables put it ahead of pi0.5 on LIBERO, RoboTwin 2.0 and RoboChallenge.

Problem it solves

Most open robot policies decide what to do from the current camera frame alone, so they lose track once an object is hidden, moved or put away, and they learn the rhythm of whoever recorded the demonstrations. Training a capable policy from scratch also needs large robot datasets few teams have. DM0.5 gives teams an open pretrained model that keeps up to a minute of visual history, was trained on several robot types, and comes with recipes to fine-tune it on their own robot data with full training or LoRA.

Key takeaways

  • Open VLA model: a 4B Gemma 3 vision-language backbone with a 680M flow-matching action expert
  • Uses key frames from up to the last 60 seconds as memory for long or partly hidden tasks
  • Generates 50-action chunks; Dexmal reports about 10 Hz on an RTX 4090 and 20 Hz on an H100
  • Checkpoints for LIBERO, RoboTwin 2.0, VLA-Arena, RoboDojo-Sim, RoboChallenge and an SO-101 task
  • Runs as the dm05 policy in LeRobot 0.7 and in Dexmal's own OpenDM code
  • Code under Apache-2.0; weights under the Gemma licence

Who it is for

  • Robot learning researchers comparing open VLA models, especially on long-horizon or memory tasks
  • Teams with bimanual ALOHA-style or single-arm setups who want a pretrained policy to fine-tune on their own data
  • LeRobot users, including SO-101 owners, who want to try a larger VLA through the usual training tools
  • Benchmark authors running LIBERO, RoboTwin 2.0 or RoboChallenge evaluations

Limitations

  • Weights are under the Gemma licence, not Apache-2.0: it allows commercial use but passes Google's prohibited-use policy and its other terms on to anyone using or sharing the model
  • All benchmark numbers and leaderboard rankings come from Dexmal or the leaderboards themselves; Dexmal publishes a technical blog, not a peer-reviewed paper
  • Training needs large GPUs: Dexmal recommends 8 GPUs, and full fine-tuning can run out of memory on an RTX 4090, where LoRA is the suggested route
  • The training data itself is not released; the blog lists its sources but not their amounts
  • The real-robot guide covers only the AgileX COBOT Magic and Dexmal's DOS-W1, with Dexmal's own camera positions
  • The base checkpoint expects a 14-value two-arm state and action layout; other robots need fine-tuning and their own normalisation statistics
  • The fast inference backend needs TensorRT, Triton and FlexAttention, and its FP8 preparation step currently has a 1,024-token limit
  • Dexmal sells a hosted version of the model, so the open release sits alongside a commercial service

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.