RESEARCH INTERFACE / STUDIO 01
Every demonstration.
One measurable scene.
Inspect the complete SO-101 tic-tac-toe corpus frame by frame, align both recorded cameras with MuJoCo, and send the same observations through the published 120K SmolVLA policy.
DATASETtic-tac-toe-so101-block-a-clean-v1
16d46485POLICYsmolvla-tic-tac-toe-games-1-15-120k
fac0cb54ENGINEMuJoCo 3.11.0 / browser WASM
250 HzINTERACTIVE / REVISION-PINNED
SO-101 Replay Studio
Dataset Replay reconstructs recorded joint state. Policy Lab asks the checkpoint for a new 50-step action chunk and executes that output in MuJoCo. The two claims never share a label.
EPISODE 000 · LOADING
MUJOCO 3.11.030 FPS DATA
Loading MuJoCo geometry
Preparing the canonical SO-101 scene…
--:--RECORDED
--:--RECORDED
MODEL EXECUTION / SMOLVLA 120K
From recorded observation to simulated motion.
Select any frame. The Studio sends its two camera views, six-axis state, and exact task to the pinned checkpoint. The returned 50 × 6 action chunk is then executed by the same MuJoCo arm—without relabeling it as a recorded success.
Checking runtimePolicy backend handshake…
GPU inference must be enabled by the Space operator. Replay remains available on CPU Basic.
50 × 6No model action has been generated yet.
Select a frame and run the revision-pinned policy.
BOARD STATEMove —
Loading task…
Board transition pending.
EPISODE METRICSrecorded
6-AXIS TRACEstate solid / action dashed
CORPUS VIEW / 195 EPISODES
Dataset diagnostics
Descriptive measurements of the recorded corpus. They are not a policy benchmark or a physical success rate.
EPISODE DURATIONseconds
TARGET COVERAGE195 moves
INTEGRITYrevision-pinned
MEASUREMENT CONTRACT / READ BEFORE INTERPRETATION
Replay is reconstruction.
Recorded state drives the arm. Contact-only kinematic calibration aligns each physical trajectory with the canonical digital scene.
Inference is new output.
Policy Lab loads the exact published checkpoint and reports its device, latency, seed, and complete 50-step action chunk.
Simulation is not hardware proof.
A simulated rollout can reveal action shape, instability, and collisions. It cannot replace bounded physical evaluation or safety gates.