Physical AI · One API

Cloud inference for every robot.

Run a robot policy as an inference endpoint on the GPU type it needs.

Backed by OPTIM NVIDIA Inception Program

The first inference platform for physical AI

Foundational infrastructure for machines that move

We run each policy on dedicated GPUs of the type it needs, with fallback cards for when one type is busy — and benchmark it automatically on simulations you build for the environment it will run in.

Run your policy

Inference API

Physical AI models behind one key, with no GPUs to spin up. Your robot policy runs as your own deployment — start from a template or bring your own checkpoint — on GPUs of its own, with workers that scale with your fleet. The vision-language and decision models it can call share the key.

  • One command to deploy. seq deploy my_policy_user — a custom policy, your model and your code, becomes a deployment in your own account, billed to it alone.
  • Straight to your worker. After one connect(), every action goes directly to your own worker over one kept-open connection — the API stays off the control loop’s hot path.
  • Scale to zero, or keep it warm. Set the minimum workers, the most GPUs and an idle window per deployment; GPU time is billed by the second, only while workers run, at least 60 seconds each time one comes up.
Test every version

Simulation benchmarks

Benchmark a policy automatically on simulations you build for its target environment — every version scored on the conditions it will face, as you iterate. Priced per simulation GPU-hour, at the same GPU rates as inference.

  • Your conditions, in simulation. Build the scenes the policy will meet — lighting, clutter, object positions — from your own environment.
  • Automatic on every checkpoint. Each version runs the same benchmark as you iterate, so a regression shows up before it reaches an arm.
  • One score to read. A pass rate per condition, with the failures grouped by what broke — not a wall of logs.

Model library

Every model and template in one library

Every model is published the same way — what it reads, what it returns, and what it costs.

Anthropicvlm
Claude Opus 5.5
claude-opus-5-5
$4.2 / $21 per Mtok · images in · claude-opus-5-5
Anthropicvlm
Claude Fable 5.1
claude-fable-5-1
$10.5 / $52.5 per Mtok · images in · claude-fable-5-1
OpenAIvlm
GPT-6 Sol
gpt-6-sol
$2.1 / $10.5 per Mtok · images in · gpt-6-sol
OpenAIvlm
GPT-6 Astra
gpt-6-astra
$10.5 / $52.5 per Mtok · images in · gpt-6-astra
Googlevlm
Gemini 3.1 Pro (preview)
gemini-3-1-pro
$2.1 / $12.6 per Mtok · images in · gemini-3-1-pro
Googlevlm
Gemini 3.8 Flash
gemini-3-8-flash
$0.7875 / $3.9375 per Mtok · images in · gemini-3-8-flash
Alibabavlm
Qwen3.8 27B
qwen3-8-27b
$0.44625 / $2.6775 per Mtok · images in · qwen3-8-27b
Perceptronvlm
Perceptron Mk1.5
perceptron-mk1-5
$0.1575 / $1.575 per Mtok · images in · perceptron-mk1-5
Anthropicvlm
Claude Haiku 5.5
claude-haiku-5-5
$0.105 / $0.525 per Mtok · images in · claude-haiku-5-5
OpenAIvlm
GPT-6 Luna
gpt-6-luna
$0.105 / $0.525 per Mtok · images in · gpt-6-luna
Z.aivlm
GLM 5.3 Flash
glm-5-3-flash
$0.1575 / $0.525 per Mtok · images in · glm-5-3-flash
Anthropicvlm
Claude Sonnet 5.5
claude-sonnet-5-5
$2.1 / $10.5 per Mtok · images in · claude-sonnet-5-5
Googlevlm
Gemini 4 Argon
gemini-4-argon
$2.1 / $10.5 per Mtok · images in · gemini-4-argon
OpenAIvlm
GPT-6.1 Sol
gpt-6-1-sol
$2.1 / $10.5 per Mtok · images in · gpt-6-1-sol
OpenAIvlm
GPT-6.1 Terra
gpt-6-1-terra
$1.05 / $5.25 per Mtok · images in · gpt-6-1-terra
Xiaomivlm
MiMo V2.6 Flash
mimo-v2-6-flash
$0.147 / $0.294 per Mtok · images in · mimo-v2-6-flash
MiniMaxvlm
MiniMax M3.1 Flash
minimax-m3-1-flash
$0.1575 / $0.63 per Mtok · images in · minimax-m3-1-flash
Moondreamvlm
Moondream 3.1
moondream-3-1
$0.315 / $2.625 per Mtok · images in · moondream-3-1
NVIDIAvlm
Nemotron 3.5 Super VL
nemotron-3-5-super-vl
$0.1575 / $0.525 per Mtok · images in · nemotron-3-5-super-vl
Alibabavlm
Qwen3.8 Omni Flash
qwen3-8-omni-flash
$0.1575 / $0.4935 per Mtok · images in · qwen3-8-omni-flash
TypeSafedecision
Jev
jev
$0.0441 / M in · structured out · jev
Physical IntelligencePolicy template
π0.5 DROID
pi05-droid
Franka Panda · 7 joint velocities · 15 Hz
seq init pi05-droid
Physical IntelligencePolicy template
π0.5 LIBERO
pi05-libero
Franka Panda · LIBERO suites
seq init pi05-libero
NVIDIAPolicy template
GR00T N1.7 DROID
gr00t-droid
Franka Panda · 7 joint positions · 15 Hz
seq init gr00t-droid
Policy template
FLUX 3 Action DROID
flux-3-action-droid
Franka Panda · 7 joints · 15 Hz · A100-40GB
seq init flux-3-action-droid
Policy template
FLUX 3 Action SO-101
flux-3-action-so101
SO-101 · 6 joints · 30 Hz · A100-40GB
seq init flux-3-action-so101
Physical IntelligencePolicy template
π0.5 DROID + Qwen
pi05-droid-qwen
Franka Panda · Qwen plans, π0.5 acts
seq init pi05-droid-qwen
Ai2Policy template
MolmoAct2 DROID
molmoact2-droid
Franka Panda · 15 Hz · L4
seq init molmoact2-droid
Ai2Policy template
MolmoAct2 YAM
molmoact2-yam
Bimanual YAM · 30 Hz · L4
seq init molmoact2-yam
Ai2Policy template
MolmoAct2 SO-100/101
molmoact2-so101
SO-100/101 · 30 Hz · L4
seq init molmoact2-so101
MicrosoftPolicy template
Rho FR3 Duo
rho-fr3-duo
Franka FR3 Duo · finetune first · 30 Hz
seq init rho-fr3-duo
Benchmark
LIBERO-Spatial
libero-spatial
10 tasks · Franka Panda
seq eval run my_policy_user --benchmark libero-spatial --adapter libero-openpi --policy-machines 2
Benchmark
LIBERO-Object
libero-object
10 tasks · Franka Panda
seq eval run my_policy_user --benchmark libero-object --adapter libero-openpi --policy-machines 2
Benchmark
LIBERO-Goal
libero-goal
10 tasks · Franka Panda
seq eval run my_policy_user --benchmark libero-goal --adapter libero-openpi --policy-machines 2
Benchmark
LIBERO-10
libero-10
10 long tasks · Franka Panda
seq eval run my_policy_user --benchmark libero-10 --adapter libero-openpi --policy-machines 2
RoboCasaBenchmark
RoboCasa365
robocasa
PandaOmron · 18 kitchen tasks
seq eval run my_policy_user --benchmark robocasa --adapter robocasa-openpi --trials 3
RoboTwinBenchmark
RoboTwin 2.0
robotwin
Aloha-AgileX · 50 dual-arm tasks
seq eval run my_policy_user --benchmark robotwin --adapter robotwin-openpi --trials 3

Start building today

Your own robot policy one seq deploy away, on dedicated GPUs — and the vision-language and decision models it can call on the same key. Nothing to provision.