Studio

Assemble the agent, then run it. Everything that decides the bill is on this page.

Brain

Model weights run on your GPU, so gradients and attention are available.

ModelAny Hugging Face VLM that fits the node.

Temperature — 0.70The notebooks run at 0.70.

Planner rate — 3.0 HzHow often the model is asked for a new command. The controller runs on regardless.

Attribution you intend to run

Saliency is generated after the drive and priced separately, but the choice belongs here: it decides which GPU the run needs. Methods your model tier cannot support say why.

Memory

What the planner carries between queries. Turning both off is a real condition, not a mistake — it is the control arm of the ablation.

Short-term visual

Recent frames — 4

Interval — 0.333 s

Send them as one native video window

As a video the model sees motion; as separate images it sees a contact sheet. The notebooks' canonical run uses four frames at 0.333 s, as video.

Long-term language

Cap — 600 characters

The planner rewrites this string itself at every query and it is truncated at the cap, so a small cap is a real constraint on what it can remember.

Body

Built inUnitree Go2 — pre-trained PPO locomotionmodel_599.pt, trained under genesis-world 1.2.0 and pinned to it. It takes a body-frame velocity command and produces joint torques. Custom agents arrive through the Agent Spec; this is the only body in v1.

Environment

Two static boxes, left and right. The canonical condition.

The notebooks' arena: grid floor, coloured boxes, a virtual back wall. Depth is ground truth, and this is the only preset comparable to published numbers.

Task prompt23/400. This text reaches a vision-language model and nothing else — never a shell, a path or an outbound request.

left_redbox, 0.4 m × 0.4 m × 0.5 m, static

right_bluebox, 0.4 m × 0.4 m × 0.5 m, static

Forward distance — 3.0 m

Lateral offset — ±1.2 m

Commit radius — 0.70 mInside this radius of a target, the trial counts as reached.

Trials

Seeds are paired and frozen across arms, so two configurations can be compared rather than just run. Free is 1 drive a day and Pro is 20; one field from one trial is noisy, which is what the convergence panel is for.

What this will cost

GPU classg5.2xlargeA10G, 22 GB, us-west-2

Drive time—Preflight plus the rollouts.

GPU cost—Spot, billed to your AWS account.

Planner calls—On your own GPU, so no per-call charge.

The drive runs first. Saliency is a separate, separately-priced step you opt into after reading the trial statistics.