ARC · VLA Playground

ARC · VLA PLAYGROUNDsimulated trajectory · target 45 FPS · measured 0 FPS

provider: canned

Embodiment

Task

Model

Provider

t = 0.00s / 0.00s (0%)

Instruction

“Pick up the red cube and place it on the white plate.”

Predicted scores

SR

—

IS

—

PS

—

Sourced from VLABench Track-1 baseline for this model.

Live phase

idle

waypoint 1 / 0

Outcome

—

Model notes

Faster decoding head; best Track-6 PS (14.4) on long-horizon composites.

How VLAs perceive · act

in wrist + 3rd-person RGB camera, language instruction, joint proprioception
out 7-DOF action (Δx, Δy, Δz, Δrpy, gripper) at 5–30 Hz
loop closed: each step the camera updates and the model re-decides — no pre-planned trajectory
note this provider replays a representative output. Toggle Live inference to call the backend.
physics contacts and grasps are scripted here for visualization. Real evaluation runs MuJoCo on the backend (VLABench-style) for actual rigid-body dynamics.