Dexbotic: A One-Stop VLA Toolbox for Pretraining, Fine-Tuning, Inference, and Evaluation

One Toolbox for the Entire VLA Pipeline

Dexbotic is a one-stop vision-language-action (VLA) toolbox open-sourced by the Dexmal team. Officially released in October 2025 with an accompanying technical report (arXiv:2510.23511), it has been updated at a high pace ever since. Instead of stitching together separate scripts for each stage, it consolidates the full VLA workflow — data preparation, pre-training, fine-tuning (full-parameter / LoRA / reinforcement learning), inference serving, simulation benchmarking, and real-robot deployment — into a single PyTorch codebase.

Three design threads run through the framework:

• Unified model interface: mainstream manipulation models (π0, π0.5, CogACT, OFT, MemVLA, GR00T N1) and navigation models (Uni-NaVid, NaVILA) share the same training, inference, and evaluation stack — switching models only requires swapping an experiment config file;
• DexData, a unified data format: official conversion scripts turn RLDS and LeRobot datasets into DexData, after which they plug directly into the training pipeline;
• Modular architecture: hierarchical configs, factory registration, and entry-point dispatch mean a new model or robot embodiment only needs to register its components — no changes to the core code.

Engineering Support for Training, Inference, and Deployment

On the training side, Dexbotic supports the FSDP2 distributed backend and ships co-training configurations for CogACT, Pi05, and DM0, letting the action expert and the language-model backbone be jointly optimized. Training is recommended from 8×A100/H100; on 8×RTX 4090 a ZeRO-3 offload config keeps memory in check. For fine-tuning, four LIBERO LoRA SFT recipes (π0, π0.5, DM0, CogACT) run even on a single consumer GPU.

On the inference side, a unified v1 inference API erases wrapper differences across models: once training finishes, the model can be served directly. With DM0, for example, you launch the inference server and send one curl request with an image and an instruction to /process_frame to get actions back. Docker is the recommended path for both development and deployment — the dexmal/dexbotic image works out of the box, and Blackwell GPUs (e.g., RTX 5090) have a dedicated c130t28 image. The minimal path:

docker pull dexmal/dexbotic
git clone https://github.com/dexmal/dexbotic && cd dexbotic
pip install -e .
torchrun --nproc_per_node=8 playground/benchmarks/libero/libero_dm0.py

DB-Pretrained Models: Consistent Gains Across Five Benchmarks

The team trained enhanced pretrained models (prefixed DB-) on their own large-scale dataset, then compared the original models against "DB-pretrained, then fine-tuned" versions on five mainstream simulation benchmarks. The gains are positive almost across the board and largest on transfer-style benchmarks like SimplerEnv and ManiSkill2 — CogACT rises from 51.25 to 69.45 and OFT from 30.23 to 76.39 — evidence that large-scale pretraining contributes far more to cross-environment generalization than in-distribution tuning.

Benchmark	Model	Original	DB-Pretrained
LIBERO (avg)	CogACT	93.6	94.9
LIBERO (avg)	MemVLA	96.7	97.0
LIBERO (avg)	GR00T N1	—	94.8
CALVIN (avg length)	CogACT	3.246	4.063
SimplerEnv (avg)	CogACT	51.25	69.45
SimplerEnv (avg)	OFT	30.23	76.39
SimplerEnv (avg)	MemVLA	71.9	84.4
ManiSkill2 (avg)	CogACT	40	58
ManiSkill2 (avg)	OFT	21	63
RoboTwin 2.0 (avg)	CogACT	43.8	58.5

DM0: A Dual-Expert VLA with Flow Matching

DM0 is the team's in-house VLA (released 2026-02-10 with its own technical report): a dual-expert structure pairing a vision-language backbone with an action expert, bridged by merged attention. Instead of discrete action tokens, the action head uses diffusion-style continuous generation via Flow Matching, producing a full chunk of future actions in a single forward pass. The open DM0-base weighs 2.4B parameters, supports 32-D actions, and accepts up to 3 camera views.

Results: DM0 averages 94.1 across the four LIBERO suites (Spatial 98.2 / Object 98.8 / Goal 96.6 / Long 82.6). On the RoboChallenge platform's Table30 real-robot evaluation, DM0 reaches a 62% average success rate and 72.25 average score, well ahead of π0 (28%) and π0.5 (43%). On the navigation front, DM0 achieves a 73.5 object-navigation success rate on HM3D and 45.3 on MP3D, on par with the navigation-specialized Uni-NaVid. On the engineering side, a Triton-backed realtime inference backend delivers roughly a 5x speedup on core inference, supporting closed-loop control on real robots.

RL Post-Training and the Real-Robot Ecosystem

On 2026-02-10, Dexbotic and the RLinf team announced a strategic partnership to jointly advance VLA + RL research. The framework has GRPO built in, and RLinf can also be used as an external RL backend for post-training, sharing the same data and model interfaces as SFT/LoRA fine-tuning.

The real-robot ecosystem already covers UR5, Franka, ALOHA, the dual-arm DOS-W1, SO-101, XLeRobot, and the Unitree G1 (including a SONIC whole-body control example), each with an inference integration tutorial. For evaluation, a separate dexbotic-benchmark repository runs the full LIBERO suite in Docker with one command, and remote real-robot evaluation can be submitted through the RoboChallenge platform. The project is MIT-licensed, model weights are hosted on Hugging Face (Dexmal), and full documentation lives at dexbotic.com.