MicroDuck Reference Target
Loading robot registry
Observation
--
Action
--
Inference
--
Algorithm
--
Platform Flow
The path from robot model to calibrated digital twin
01
Robot Registry
Observation/action contract, simulator, deployment format
02
Training Job
MuJoCo plus PPO worker boundary
03
Policy Registry
ONNX artifact plus metrics manifest
04
Sim-to-Real Calibration
Telemetry mismatch and actuator parameter estimates
Emulator Contract View
Visualizes the current MicroDuck interface before the real MuJoCo runtime is attached
Digital Twin Operations
3D state linked
asset loading
phase --
Action bars
Velocity
Body height
Mode
--
Cadence
--
Battery
--
Gait phase
--
Navigation Map
intent standby
synthetic vision
x 0.0 / y 0.0
heading 0
target docking
Interface Contract
| Field | Value | Purpose |
|---|---|---|
| observation | 61 | Runtime input vector accepted by policy. |
| proprioception | 48 | Joint, IMU, and body-state feedback. |
| command | 13 | Operator or task target conditioning. |
| action | 14 | Controlled joint output vector. |
| policy | ONNX / 50 Hz | Deployment artifact and inference cadence. |
Current state: this is a contract visualization, not a live emulator. Next integration step is attaching MuJoCo state frames and policy rollout snapshots to this panel.
Mission Control
Walk validation
Advance gait frames and verify policy output stability.
Stability pulse
Step a short rollout and watch base height response.
Interface check
Confirm observation, action, and telemetry contracts.
Morphology
Control Patterns
Direction Control
Demo readiness
Synthetic runtime, policy contract, and local inputs are available.
Simulator Runtime
loading
Step
--
Time
--
Velocity
--
Height
--
--
Integration Readiness 2/4
Control plane
Config, API, GUI, artifacts
ready
Policy contract
Obs/action shape validation
ready
MuJoCo runtime
State frames and rollout loop
pending
Hardware telemetry
Real robot trace ingestion
future validation
Browser Interaction
Optional local camera and mic inputs that emit simulator commands
local only
Browser Interaction
Optional local camera and mic inputs that emit simulator commands
Camera off
Camera input
Preview only in this first pass
off
Mic command
Say reset, step, walk, or stop
off
Waiting for local input.
Planning
Boundaries, components, and phase implementation plan
Work Boundary
Workable now
No hardware required
- Robot and task registry
- Observation/action validation
- ONNX policy registry shape
- Training/eval job API contract
Simulated substitute
Useful before hardware exists
- Synthetic telemetry traces
- Delay, noise, voltage, friction perturbations
- Policy rollout viewer
- Calibration dry run against simulated reality gap
Hardware later
Validation target, not a blocker
- Real robot telemetry ingestion
- Motor response measurement
- Command latency measurement
- Policy deployment validation
Component Design
Registry
Robot, task, policy metadata
Simulation
MuJoCo model, rollout frames
Training
PPO jobs and metrics
Telemetry
Synthetic now, hardware later
Phase Plan
P0
Control-plane MVP
GUI, config, artifacts, smoke jobs
P1
Simulation runtime
MuJoCo import, rollout snapshots, viewer
P2
Training worker
PPO jobs, metrics, ONNX export
P3
Hardware validation
Replay logs, real telemetry, deployment checks
Experiment Manager
Loading experiment plan
Task
--
Parallel Envs
--
Steps
--
Worker Log
Responses from web-api dispatch routes
Ready.
Policy Registry
Generated ONNX manifests and local artifacts
Telemetry and Digital Twin
Reality-gap hooks for joint, IMU, delay, voltage, and actuator calibration
Joint Compare
ready
Delay Estimate
ready
Twin Update
next
Runtime
Container configuration and service state
Loading runtime