How dexmal/opendm works
OpenDM is a vision-language-action model toolkit for training and deploying DM0.5 across simulated and physical robot embodiments. The main workflow is to configure a task recipe, load and transform robot demonstrations, fine-tune the model, then serve policy inference for robot observations. The graph also includes dataset registration, evaluation-oriented recipes, and the optimized TensorRT inference path. Relationships among sampled modules follow the source index or excerpts; broader evaluation and deployment capabilities are documented in the README, while unsampled implementation details are not inferred.
Main entry points: dm05_exp.py DM05 Experiment · transforms.py Data Transforms · dm05_infer.py Inference Service · dm05_robotwin2.py Task Recipe
Components 21
Task Recipes
- Task Recipe
playground/dm05_robotwin2.py - Evaluation Recipes
playground
Robot Data
- Dataset Registry
opendm/dataset/register.py - Robot Datasets
opendm/dataset - JSONL Dataset
opendm/data/dataset.py - Data Transforms
opendm/data/transforms.py - State Normalization
opendm/data/normalize.py - Image Augmentation
opendm/data/augmentations.py - Batch Collator
opendm/data/collator.py - Robot Specifications
opendm/constants/robot.py
Model Training
- DM05 Experiment
opendm/exp/dm05_exp.py - DM05 Model
opendm/model/dm05/dm05_arch.py - LoRA Fine-tuning
opendm/model/dm05/dm05_lora.py - Optimizer
opendm/optimizer/muon_adamw.py - Trainer
opendm/trainer/trainer.py
Policy Runtime
- Inference Service
opendm/infer/dm05_infer.py - Fast Decode
opendm/infer/dm05_infer_arch.py - TensorRT Vision
opendm/infer/build_vision_trt.py - FP8 Prefix Backend
opendm/infer/dm05_prefix_mlp_trt.py
Other components
- Researcher
- Robot Client
Connections 22
- Researcher → Task Recipe: configures
- Researcher → Evaluation Recipes: evaluates
- Task Recipe → DM05 Experiment: configures
playground/dm05_robotwin2.py - Evaluation Recipes → DM05 Experiment: configures
playground/dm05_vla_arena.py - Dataset Registry → Robot Datasets: imports
opendm/dataset/register.py - Task Recipe → Dataset Registry: registers
playground/dm05_robotwin2.py - DM05 Experiment → Dataset Registry: uses
opendm/exp/dm05_exp.py - DM05 Experiment → JSONL Dataset: uses
opendm/exp/dm05_exp.py - DM05 Experiment → Data Transforms: builds inputs with
opendm/exp/dm05_exp.py - DM05 Experiment → Batch Collator: batches with
opendm/exp/dm05_exp.py - Data Transforms → State Normalization: uses
opendm/data/transforms.py - Data Transforms → Image Augmentation: uses
opendm/data/transforms.py - Data Transforms → Robot Specifications: uses
opendm/data/transforms.py - DM05 Experiment → DM05 Model: configures
opendm/exp/dm05_exp.py - DM05 Experiment → LoRA Fine-tuning: uses
opendm/exp/dm05_exp.py - DM05 Experiment → Optimizer: configures
opendm/exp/dm05_exp.py - DM05 Experiment → Trainer: runs
opendm/exp/dm05_exp.py - Robot Client → Inference Service: sends observations
- Inference Service → Fast Decode: dispatches decode
opendm/infer/dm05_infer.py - Inference Service → TensorRT Vision: runs vision engine
opendm/infer/dm05_infer.py - Inference Service → FP8 Prefix Backend: optionally uses
opendm/infer/dm05_infer.py - Fast Decode → DM05 Model: uses model
opendm/infer/dm05_infer_arch.py