Files
VibemakerSandbox/app/public/robots/apollo/SOURCE.md
T

4.1 KiB

Apptronik Apollo

Source and checkpoint

  • Model: /home/cen/Embodied_Workspace/PlatformProject/apptronik_apollo/apptronik_apollo.xml and its STL assets (Apache-2.0, copyright Apptronik; see LICENSES/apollo.txt).
  • Training: /home/cen/Embodied_Workspace/Mujoco_Projects/mjlab (see LICENSES/mjlab.txt).
  • Explicitly selected v0.6 checkpoint: logs/rsl_rl/apollo_flat_upright_v3/2026-11-02_17-39-21_apollo_upright_v3/model_9999.pt.
  • Browser policy: corresponding run's ONNX, copied to policies/walk.onnx. Import verifies exported MLP weights, observation mean and folded normalization standard deviation against the checkpoint, not just its file timestamp.
  • source-manifest.json records checkpoint SHA-256, copied file hashes and training configuration hashes. No runtime access to mjlab is required.

Contract and modifications

The 108 observations are body-frame linear velocity (base IMU velocimeter), angular velocity (gyro), projected gravity, 32 relative joint positions, 32 joint velocities, 32 previous actions and the three-component velocity command, in that order. Normalization is embedded in ONNX. Action targets are stand-pose defaults plus per-joint residual scales. Head/arm residual scales are zero, as in the training configuration; their original actuators hold the pose.

import-apollo.py applies mjlab Entity's collision editors and preserves native actuator gains, damping, armature, force limits and stand defaults. Generated model.xml is modified from upstream: named visual geoms, base-aligned IMU, foot sites, training collision filtering, no standalone lights/keyframes, explicit task solver options (5 ms timestep, four substeps per control tick, Newton/pyramidal/implicitfast, 20 iterations). The newer native MuJoCo serializer's input="pos" actuator keyword is omitted for WASM compatibility; the equivalent general actuator's position gain/bias is retained. Stand defaults are read from mjlab's calibrated upright-v3 keyframe, not the original deeply flexed asset pose or rounded ONNX metadata. The importer cross-checks the pose against ONNX metadata. Knees default to 0.35 rad, spawn root height is 1.1008385 m (15 mm clearance), grounded target height is 1.0858385 m, and the fall threshold is 0.98 m, matching upright-v3 training.

The shared platform composer creates the floor with explicit collision bits and group, overriding Apollo's visual-only global geom defaults. The arena uses the platform's normal robot-height scaling, interactive objects and camera controls. No pose animation or kinematic walking is used.

Controls: W/S = +0.5/-0.3 m/s forward/backward; A/D = +/-0.4 rad/s yaw command; arrows are aliases. Key release zeros the command. Space resets; Escape pauses; C restores chase camera.

Re-import

From the mjlab directory:

uv run python /home/cen/Embodied_Workspace/Mujoco_Projects/microduck-simulator/app/scripts/import-apollo.py . \
  --checkpoint-file logs/rsl_rl/apollo_flat_upright_v3/2026-11-02_17-39-21_apollo_upright_v3/model_9999.pt

Use --checkpoint-file to reproduce this release exactly. Without it, the importer selects the newest checkpoint in apollo_flat_upright_v3 by modification time (iteration breaks ties). Its run must already contain a matching ONNX export newer than the checkpoint; otherwise import fails rather than silently using stale weights. Set APOLLO_ASSETS_ROOT if the source model lives elsewhere.

Verification

cd app
npm run validate:packages
npx playwright test apollo.spec.js

Real browser MuJoCo WASM and ONNX inference test: 108 observations / 32 actions, WASD commands and key release, floor collision bits, five independent five-second finite-state rollouts, reset, and switching between Apollo/T1/RP1/Microban. Upright-v3 rollouts with props parked: no falls; forward x +2.258 m, backward x -1.174 m, left yaw +1.158 rad, right yaw -1.725 rad, standing height 1.086 m. Tests also require final root height above 1.04 m in every trial. This is browser simulation acceptance, not a claim of perfect velocity tracking or hardware equivalence. Screenshot is generated in Playwright's test-results directory.