Go2 world-model results
All measured comparisons in the paper’s scaling, transfer and Go2 replay panels, with scratch-pilot logs and original run diagnostics. Snapshot through the 18:23:59 TrotRun recording on 22 November 2026.
Scan-conditioned scaling in simulation
Five widths, four exposure budgets, one training seed per width. The 20 checkpoints share 24,576 evaluation windows in 96 held-out shards. Exposure counts sampled targets including repeats. One-step NRMSE scores the change in the observation. The longer horizons score the observation at the specified endpoint.
Fitted NRMSE = E + A n⁻ᵅ + B d⁻ᵝ. Size n is measured in 30M parameters and exposure d in billions.
The last exposure doubling at 120M improved three-second error on 92/96 shards (absolute decrease 95% paired interval 0.018–0.056). At comparable estimated compute, 16M long versus 120M short reduced error by 21.4% on the pooled panel and won on 96/96 shards (absolute decrease interval 0.080–0.132). Corrected estimated FLOPs were 5.42 × 10¹⁷ and 5.21 × 10¹⁷. The unstable 8M, 0.721B branch produced the largest discrepancy at every horizon.
Withholding the final doubling
Mean absolute relative discrepancy was 2.9% for the extrapolated fit and 7.4% for carrying the preceding measurement forward. On the three-second fit, an alternate solver gave (E, α, β) = (0.178, 1.186, 0.222). Dropping the unstable branch gave (0.233, 0.499, 0.289). Also correcting exposure for surviving trunk updates drove E to its lower bound and gave (α, β) = (0.051, 0.253). Leaving out one width at a time yielded α between 0.035 and 2.
Removing the terrain scan
The 120M parent becomes a 119,670,302-parameter sensor-only model. The update changes 749,568 weights and biases in the input projection using 2,048 sampled simulated target transitions. The four endpoint forecasts below share 32 held-out flat worlds and a common normalization.
Adapted minus initial paired 95% intervals span −0.402 to −0.027 at 1 s and −0.322 to 0.439 at 3 s. Against persistence the corresponding intervals are −0.711 to −0.395 and −0.502 to 0.059.
All five sizes and scratch training pilots
The five projection experiments use 64 training worlds and 32 evaluation worlds. The scratch pilots use width 128, depth 3, 8 or 64 training worlds, 64 validation worlds and 600 updates per cohort. The logs retain each experiment’s normalization and sampled evaluation cohort.
Physical Go2 and simulation replay inventory
Replay identifiers below match the supplement. The 20 zero-scan models come from the same 20 scan-conditioned checkpoints in the scaling sweep. In recorded cases the physical controller generated the published joint targets once. The predictors forecast the recorded 30-channel sensor stream open-loop at 50 Hz.
Learned outputs are 12 joint angles, 12 joint velocities, three body angular rates and the three-component body-up vector. MuJoCo integrates joint state and floating-base dynamics and derives corresponding sensor signals. Global position and linear speed are absent from the learned output. SDK SportClient and the simulated T2 policy are command sources.
Pooled three-second physical replay
Nine stock-SDK recordings on one Go2. Four have a documented 0.5 m/s request. The two handset recordings form a separate panel. Each trial contributes 150 forecast ticks. Equal-length joint and gyro RMSEs pool in quadrature. Up error averages frame-wise angles. Lower values indicate better predictions.
Default SDK, all nine
Default SDK, four at 0.5 m/s
Handset, two recordings
Two shorter physical SDK requests
Both requested 0.5 m/s and a 0.45 m goal. StaticWalk reached 0.459 m of controller odometry in 1.308 s. TrotRun timed out at 0.362 m in 2.013 s. The common 1.00 s window includes the mode transition. The reported gait code stayed zero. Each request has one recording.
Eight T2 gaits in MuJoCo
One three-second trace per request at 0.5 m/s, on flat ground from a common standing initialization. The reference is MuJoCo driven by the T2 policy. The pronk reference overturned. Each predictor receives the joint targets for its corresponding trace. The physical SDK and T2 policy generated different trajectories.
Reference trajectory
Filmed 2 m trial and repeatability
For 20261122-180625-ueuixi21, the full 5.08 s replay uses SDK-generated targets and the recording of that same robot. The three forecast rows below are the complete walk. The four repeat runs each provide a matched three-second comparison.
Four stock SDK runs at 0.5 m/s
The two-foot grid and four stationary contacts gave 2.335 m of stance-centroid travel. Independent picks and refit gave 2.344 m. A two-pixel coherent shift gave 2.286–2.397 m, a sensitivity range. Estimated camera travel exceeded settled base odometry by about 0.30 m on this one walk. Motion continued briefly after the stop command. Odometry settled within 1 cm by +0.12 s. Average stop-window odometry speed was 0.394 m/s. Camera travel divided by the visible-motion interval gave about 0.47 m/s (0.44–0.48 m/s with alternate pixel thresholds). Apparent first visible motion followed the mapped command by 0.284 s. The clock estimate has a 47 ms SSH half round trip and camera transport delay is unmeasured.
Every fixed three-second comparison
Each case opens a table of all scored predictors. There are 37 predictor rows per physical case and 35 per simulated gait. In simulation MuJoCo is the reference. Core and adapted supplement tables supply 687 distinct case–predictor pairs.
Default SDK · 20261122-172206-4hzkt4or
Default SDK · 20261122-173234-nj1x4r2n
Default SDK · 20261122-173709-6kcsby0i
Default SDK · 20261122-174128-ozhk4ftt
Default SDK · 20261122-175237-njtx30s_
Default SDK · 20261122-175321-do0cvben
Default SDK · 20261122-175903-x4su5g5f
Default SDK · 20261122-180247-qg3m_9bx
Default SDK · 20261122-180625-ueuixi21
Handset · 20261122-174446-r9_dra2k
Handset · 20261122-174838-yxr7pq_2
T2 simulation · walk
T2 simulation · trot
T2 simulation · flying_trot
T2 simulation · pace
T2 simulation · bound
T2 simulation · pronk
T2 simulation · crouch_walk
T2 simulation · stair_climb
Original per-run diagnostics
These additional rows preserve the original scoring window for each run, including repeated simulation rehearsals. Window lengths and controller identities vary. A physical score requires a recorded reference. A rehearsal can be present even when a physical trial did not complete.