State-of-the-art policies on managed GPUs.
Launch a training run in one call, from lightweight baselines to billion-parameter VLAs, on cloud GPUs Neuracore provisions, schedules, and fails over for you. Or bring your own model. No infrastructure to run.
From baselines to billion-parameter VLAs.
Each ships with a tuned Hydra config, outputs action chunks of length output_prediction_horizon, and supports cross-embodiment training with visual-only or proprio-only configs via masking.
ACT
Action Chunking with Transformers, encoder-decoder with a CVAE latent, predicting action chunks from images + proprioception.
CNN-MLP
Per-camera CNN encoders + multimodal fusion + MLP head. The baseline with the broadest modality support.
Diffusion Policy
Visuomotor policy via action diffusion, 1D conditional U-Net with FiLM conditioning, plus a first-class flow-matching mode.
π0 · π0.5
Vision-Language-Action flow-matching models, PaliGemma + Gemma action expert; π0.5 adds open-world generalization and quantile normalization.
GR00T N1.6
NVIDIA GR00T VLA, Cosmos-Reason/Eagle VLM + a 32-layer Diffusion Transformer action head with rectified flow matching.
Custom
Subclass NeuracoreModel with forward / training_step / configure_optimizers and declare your input/output DataTypes.
Kick off a GPU run in one line.
Neuracore provisions a Deep-Learning PyTorch VM, installs the package, and runs the Hydra entrypoint. Live regional quota is checked before launch and the job fails over across zones on resource exhaustion.
Set batch_size="auto" and an affine GPU-memory model fits the batch that fills ~90% of VRAM, downshifting on OOM.
job = nc.start_training_run(
name="cube-handover-run",
dataset_name="Cube Handover",
algorithm_name="DiffusionPolicy",
gpu_type="A100_80GB",
num_gpus=4,
frequency=10,
algorithm_config={
"batch_size": "auto",
"output_prediction_horizon": 16,
},
)
# Resume later from the latest checkpoint
nc.resume_training_run(job.id, additional_epochs=50)Your model, or ours, same managed stack.
Bring a custom algorithm and Neuracore validates it on a real GPU before it ever trains; or fine-tune a foundation VLA with per-component freeze controls.
Bring your own model, validated on a GPU
Uploads are AST-validated, hyperparameters auto-extracted into UI fields, then a T4 VM runs a real forward/backward/optimizer/export/endpoint checklist before the algorithm is marked available.
Fine-tune foundation VLAs
Per-component freeze/unfreeze (action expert, vision encoder, top-N LLM layers), differential LR scales, gradient checkpointing, torch.compile, and bfloat16.
Cross-embodiment data loading
PytorchSynchronizedDataset caches episodes, computes dataset statistics, and zero-fills missing sensors with per-sensor masks so mixed-robot batches train correctly.
Monitoring & checkpoints
TensorBoard scalars, images, and gradient/weight histograms; cloud metrics for CPU/GPU/memory; periodic checkpoints with last-N retention and resumable restore.
GPU-aware scheduling checks live quota before launch. DistributedDataParallel over NCCL, batch-size autotuning, checkpoint/resume, and TensorBoard + cloud metrics come standard.