How dexmal/dexbotic works
Dexbotic is a PyTorch toolbox for developing vision-language-action systems: users prepare robot or simulation data, train or fine-tune a selected policy, then run inference and evaluate or deploy it on a robot. The tree documents a broad family of manipulation and navigation models, plus supervised, co-training, and RL paths. The excerpts directly show multimodal model construction, tokenization, an inference client, and RL rollout/training behavior; connections for other listed subsystems reflect README-described capabilities and the repository layout, not inspected implementations.
Main entry points: exp Experiment configs · rl_trainer.py RL trainer · model VLA model families · client.py Robot inference client
Components 20
Data preparation
- Dataset adapters
dexbotic/data/data_source - Dataset pipeline
dexbotic/data/dataset - Batch collation
dexbotic/data/collator.py - Conversation tokens
dexbotic/tokenization/process.py
Training
- Experiment configs
dexbotic/exp - Training engine
dexbotic/exp/trainer.py
VLA models
- Multimodal core
dexbotic/model/dexbotic_arch.py - VLA model families
dexbotic/model
Inference deployment
- Policy wrappers
dexbotic/policy - Inference runtime
dexbotic/infer/runtime.py - Robot inference client
dexbotic/client.py - Robot bridges
hardware - Simulation environments
dexbotic/sim_envs - Policy evaluation
dexbotic/exp
RL post-training
- RL trainer
dexbotic/exp/rl/rl_trainer.py - RL rollouts
dexbotic/exp/rl/rl_rollout_redis.py - RL backend bridge
dexbotic/rl/rlinf_bridge
Other components
- Researcher
- Robot operator
- Robot datasets
Connections 22
- Researcher → Experiment configs: configures runs
- Robot datasets → Dataset adapters: loads from
- Dataset adapters → Dataset pipeline: provides records
- Dataset pipeline → Batch collation: prepares samples
- Batch collation → Conversation tokens: batches text
- Experiment configs → Training engine: configures training
- Training engine → VLA model families: trains
- VLA model families → Multimodal core: uses multimodal core
- Experiment configs → Dataset pipeline: selects data
- Robot operator → Robot inference client: sends observations
- Robot inference client → Inference runtime: requests predictions
- Inference runtime → Policy wrappers: dispatches inference
- Policy wrappers → VLA model families: runs model
- Robot inference client → Robot operator: returns actions
- Robot bridges → Inference runtime: bridges robot I/O
- Simulation environments → Policy evaluation: provides episodes
- Policy evaluation → Policy wrappers: evaluates
- Experiment configs → RL trainer: configures RL
- RL trainer → RL rollouts: collects trajectories
- RL rollouts → Simulation environments: steps environments
- RL trainer → VLA model families: updates policy
- RL backend bridge → RL trainer: integrates backend