VLA Foundry: A Unified Framework for Training Vision-Language-Action Models

Figure 2 : Throughput scaling as the number of GPU nodes is increased for the LLM, VLM, and VLA with either DDP or FSDP parallelization. Tests were done through Sagemaker on P5 nodes of 8 Nvidia H100 GPUs each. See section 4 for details about the models.

Figure 3 : VLM 1.1B caption-only model predictions (greedy decoding). The model uses normalized, 224 \times 224 input images to generate the captions. Images were sampled from the authors’ phones (+ logo) to avoid any contamination.

Figure 4 : Foundry-VLA-1.7B architecture. Four images over two timesteps each are fed into the same ViT image encoder. For each of the 8 images, the result is pooled with “pixel-shuffle” [ marafioti2025smolvlm ] (see appendix 8.4 ) and projected into the embedding space of the LLM. An additional observation token is appended to the sequence. The LLM embedding of the last layer matching this token is passed to a flow transformer with a noised action sequence. The flow transformer outputs the predicted denoising direction.

Figure 6 : Overview of seen simulation evaluation tasks. The lbm_eval_oss task suite spans tasks that require different qualities of manipulation capabilities: from pick-and-place to non-prehensile manipulation to bimanual coordination. Here, we show a single still from about the midpoint of a successful rollout from Foundry-Qwen3VLA-2.1B-MT . Video versions of these images can be found at https://tri-ml.github.io/vla_foundry . We note that the images here build on top of [ pfaff2025drakeblendertools ] , where we re-light and re-render the Meshcat scenes at the desired frame rate from rollouts

Figure 1 : VLA Foundry overview. Unified and Configurable LLM-VLM-VLA Pipeline : VLA Foundry was designed to enable flexible model composition. For example, users can train an LLM, use the LLM to train a VLM, and use the VLM to train a VLA. Bootstrap off Pre-trained Models : VLA Foundry also supports loading off-the-shelf VLMs from Hugging Face. Pre-trained Open Models : We release LLM, VLM, and VLA models trained both from scratch and finetuned from open weights under a permissive license at https://huggingface.co/collections/TRI-ML/vla-foundry .

Figure 5 : We compare our multi-task models— Foundry-VLA-1.7B-MT-sim , Foundry-VLA-1.7B - full , and Foundry-Qwen3VLA-2.1B-MT —against the LBM-MT [ lbmtri2025 ] multi-task model on a set of seen tasks in lbm_eval_cs . In aggregate, LBM-MT and Foundry-VLA-1.7B-MT-sim are on par, while Foundry-Qwen3VLA-2.1B-MT far outperforms the rest. Note that here only Foundry-VLA-1.7B - full and Foundry-Qwen3VLA-2.1B-MT share the same exact robot training data; for more details refer to Section 4.3.1 .

Figure 9 : Simulation results of our three multitask Foundry-VLA-1.7B variants: trained on simulated data Foundry-VLA-1.7B-MT-sim , real data Foundry-VLA-1.7B-MT-real , and both combined Foundry-VLA-1.7B-MT .