Robotics / VLA

Most VLA models on this page are research-only — Apache 2.0 or MIT license on the code, restricted to non-commercial research on the weights. A few (π-0-fast, RDT-1B) ship with broader licenses. Always read the model card before deploying on a paid product. Commercial managed-API offerings are expected to roll out across 2026.

Inference typically requires a single H100 or A100 GPU per robot at 10-50 Hz. Smaller distilled policies (Octo-small, π-0-fast) can run on a single 4090 or A6000. For research, most labs run them on workstations adjacent to the robot. For production, expect to dedicate one GPU per active robot or one shared GPU across a small fleet.

On shared-GPU platforms (Vast.ai, Replicate) you pay per second of compute; the cost per inference step depends on the GPU and the model size. Self-hosted on your own hardware is electricity plus depreciation. On Railwail, the VLA models are not currently available to run.

Most VLAs are trained on specific platforms — Franka Panda, UR5, ALOHA, mobile ALOHA, Cobot Magic, etc. Cross-embodiment generalization is improving (Octo and RT-X were explicit attempts) but deploying on a new arm still typically requires 100-1,000 fine-tuning demonstrations. Check the model card for trained embodiments.

Tabletop pick-and-place is reliable on most VLAs. Multi-finger dexterity, in-hand manipulation, and tool use are still hard — they work in demos but generalize poorly. π-0 and RT-2 show the strongest dexterity to date in open research; expect rapid progress through 2026.

A regular policy maps observations to actions. A VLA additionally conditions on a natural-language instruction, so the same policy can do 'pick up the red mug' and 'pick up the blue cup' from the same model. This shifts complexity from per-task training to large-scale instruction-action pretraining.

Collect 100-1,000 teleoperated demonstrations of your target tasks, then run supervised fine-tuning (typically LoRA) on the pretrained checkpoint. Most repositories include a fine-tuning script. Plan for 4-24 hours of GPU time per fine-tune on a single H100, plus a few days of evaluation iteration.

Physical Intelligence, Skild AI, Covariant, and a handful of stealth labs are explicitly building general-purpose commercial VLAs with managed APIs. Expect the first commercial offerings (likely vertically integrated with specific robot OEMs) to ship across 2026 and 2027. Railwail will list them here as they launch.