Baidu's LoongForge-Embodied puts nine heterogeneous embodied models on one training stack, lifting end-to-end throughput up to 4.38x
Synopsis
Baidu released LoongForge-Embodied, a native-PyTorch training framework that lets vision-language-action models and world-action models, two structurally different families of embodied models, share one training infrastructure. Previously these models each maintained separate training stacks, so optimizations could not be reused across them. The framework separates model-specific computation from reusable infrastructure and reports 1.47x to 4.38x end-to-end training throughput gains over each model's open-source implementation across nine models, with benchmark success rates reported as comparable to official baselines.
Interpretation
Training embodied foundation models no longer requires a separate stack per model: vision-language-action models and world-action models can share training orchestration, distributed execution, data processing, checkpoint management and evaluation inside one framework. Many embodied foundation models previously relied on independently developed training stacks with customized distributed execution, data processing and evaluation pipelines, so system techniques such as optimized attention, graph compilation, CUDA Graphs and communication optimization had to be integrated, tuned and validated repeatedly across separate stacks. Across nine representative models, end-to-end training throughput improves by 1.47x to 4.38x relative to each model's corresponding open-source implementation; DreamZero (Wan2.2-TI2V-5B) shows the largest gain at 4.38x and Wall-OSS-0.5 the smallest at 1.47x.
The framework separates model-specific computation from reusable training infrastructure, so heterogeneous models keep their own computation graphs and training objectives while sharing common system-level capabilities. Configuration-driven model registration, a lightweight model interface and an extensible training engine form the unified abstraction layer; a model implements construction and one forward operation that returns a loss and diagnostic metrics. Evaluation covers nine models, including vision-language-action models, world-action models and the omnimodal world model Cosmos 3, with success rates reported on the LIBERO, RoboTwin and SimplerEnv simulation benchmarks.
The same abstraction extends to evaluation, where training and evaluation reuse one model inference implementation and the evaluation loop is accelerated through concurrency, lower per-episode cost and reuse of policy-independent computation. The reference path runs eight independent client-simulator pairs, one per GPU, without a shared coordinator; the accelerated path drives the same eight GPUs through a larger slot pool with the same checkpoint, benchmark protocol and renderer. Across six benchmark runs, end-to-end evaluation time speeds up by 4.1x to 23.2x; a full 5,000-episode RoboTwin run drops from 41.03 to 1.77 hours, and a 2,000-episode LIBERO run under software rendering drops from 173.7 to 16.3 minutes.
Perspective
The demonstrated scope is training throughput on nine embodied foundation models and downstream success rates on the LIBERO, RoboTwin and SimplerEnv simulation benchmarks; evaluation acceleration is measured on an eight-GPU configuration across six benchmark runs, with the reference and accelerated paths using the same checkpoint and benchmark protocol.
The training-fidelity loss comparison is visual and limited to the displayed training windows, so it does not by itself establish full-run convergence or numerical identity; the low-precision communication paths report small deviations only within short training windows and do not establish long-horizon convergence or downstream task equivalence. Accelerated evaluation is judged by aggregate success rates under the same protocol rather than by episode-by-episode reproduction.
