orbit the robot. change its checkpoint. inspect the physics, then follow the evidence below.
live checkpoint explorer · collapse to pause
checkpoint explorer
inspect the checkpoints
Loading MuJoCo WebAssembly…
walk_fast_500 runs the saved mean-policy network and gait generator locally. switching checkpoints creates a fresh physics state.
20 s history · time →
Torso xForward vRoll / pitch
Drag to orbit. Right-drag or use two fingers to pan. Scroll or pinch to zoom. Follow C-1N tracks the live torso until you move the camera yourself.
Current model mass
Loading…
Gravity
Loading…
Loading current force pulse…
Simulation time
—
Torso x
—
Torso y
—
Torso height
—
Contacts
—
Roll / pitch
—
Joint-speed RMS
—
Actuator-force RMS
—
STAND force-pulse suite
33 declared external-disturbance cases
STAND loops through all 33 cases by default. Each nonzero case applies a 200 ms horizontal force pulse at reset, then gets 1 second of observation before the next case. Pause the suite or use the compact controls to inspect one case directly.
Loading perturbation case…
Coverage map: one control and four force-magnitude rows across eight bearings. The highlight shows the case now in view; the map is not interactive.
The eight declared 1 mg force-pulse cases are the current recovery frontier; the canonical record retains 1 mg recovery as failed. Browser playback uses the canonical STAND model, reset, and 33-case definition. Its simplified visual controller is not a parity port of native contact/support-margin gating or Jacobian-transpose control.
Live telemetry
Signals from the selected checkpoint
Each graph keeps the latest 20 seconds of MuJoCo state. The horizontal axis is time. Reset or checkpoint selection clears the sample history.
Ground-plane positionx y · metersTorso heightmetersBody attituderoll pitch · radians
Runtime: first-party MuJoCo JavaScript bindings ↗, pinned to v3.11.0. The STAND model, deterministic reset, 33-case force-pulse definition, and Googly Surveyor visual layer are pinned to C-1N commit a6cf1cb ↗. The browser visualizes a simplified bounded controller; the native rollout remains the evidence record.
c-1n // 03 · stride
watch the walk change.
four recorded policies. the same five seconds, played at normal speed. a second camera shows the same trajectories from the side.
mean-policy speed
0.888 m/s
fixed evaluations
0 / 24 falls
training checkpoint
500 updates
the fastest checkpoint is 2.80× the stable walk in this recorded mean-policy run. speed and stability are measured; foot slip and contact fragmentation remain.
PPO first learns joint corrections around a tripod reference. the faster policy also controls gait cadence. the chassis stays six-legged; the checkpoint preserves the settings, weights, and measured runs.
five-second mean-policy recordings · seed 201
checkpoint
speed
role
notebook baseline
0.379 m/s
original chassis / crude PPO
walk_stable_100
0.318 m/s
locked walking baseline
walk_fast_250
0.702 m/s
intermediate cadence policy
walk_fast_500
0.888 m/s
accepted STRIDE checkpoint
the chassis and action space change between the notebook and walking policies. this comparison shows the recorded progression, not a controlled test of training duration alone.
walk_fast_500 averages 0.821 m/s in sampled evaluation, with no falls in the 24 fixed flat-ground evaluations. deterministic mean runs share an initial state and are not independent robustness trials. terrain, push recovery, and energy efficiency are not established.
Checkpoint lineage
The robot should make its own evolution inspectable.
C-1N is the integration point. The Robotics Test Bench isolates the physical questions that make a checkpoint legible. A checkpoint advances only for a demonstrated capability or an understood failure worth preserving and comparing.
C-1N // 00 · SPAWN — historical deterministic six-foot spawn baseline. It does not claim stance stabilization or locomotion.
C-1N // 01 · SHUFFLE — preserved coordinated-gait failure. Motion is not controlled locomotion.
C-1N // 02 · STAND — reproducible stationary-stance baseline. Its closed-loop controller maps Cartesian foot-position error into bounded joint-space corrections through a Jacobian transpose, gated by declared six-foot contact and positive COM-projection margin. This is not robust disturbance rejection.
C-1N // 03 · STRIDE — completed by walk_fast_500. The learned cadence policy is preserved with fixed flat-ground evaluations and the recorded comparison above. Foot slip and contact fragmentation remain; terrain and push robustness are unproven.