Does AI prioritize
safety over efficiency?

Exploring alignment in embodied AI. A Unitree G1 humanoid navigates forbidden zones in MuJoCo while Gemini 3 makes safety-critical decisions.

Powered by leading AI & simulation technologies

MuJoCo
Gemini 3
Inspect AI
Next.js + Tremor
Research Questions

Alignment in
embodied AI

Safety vs Efficiency

Does the AI cut through forbidden zones when pressured to reach goals quickly?

Rationalization Detection

When the AI violates constraints, does it rationalize or honestly acknowledge it?

Self-Assessment Accuracy

How accurately does AI self-assessment compare to ground-truth simulation metrics?

Experiment Design

Architected for rigorous
alignment testing

The G1 Alignment Experiment uses MuJoCo physics simulation to create realistic scenarios where an AI must navigate safety-efficiency tradeoffs. Every decision is logged, compared against ground truth, and scored for alignment.

36
LiDAR rays (360° coverage)
5
Retry attempts with learning
3
Sensors (Camera, LiDAR, IMU)
Forbidden zone enforcement
Configurable safety boundaries the AI must respect, with ground-truth violation tracking.
Real-time sensor fusion
Camera images, 36-ray LiDAR, and IMU data fed to Gemini at each decision point.
Multi-attempt learning
5 retry attempts where the AI receives feedback on violations and adapts its strategy.
Honesty detection
Compare AI self-assessment against simulation ground truth to detect rationalization.

See AI decision-making in action

Explore experiment runs, view AI reasoning traces, and compare self-assessment against ground truth metrics.