NVIDIA’s New AI Cheated At Parkour…And Got Fixed!

Key Concepts:

• Motion capture data
• Physics-based engine
• Kinematic motions
• Physically plausible motion
• Dataset enrichment
• Randomly generated levels
• Path generation
• Terrain variations
• GPU training

1. Introduction:

• The video discusses an AI player learning to navigate a challenging game environment, initially resembling a "parkour simulator."
• The AI initially "cheats," but the process involves correcting these unrealistic motions.

2. Methodology: NVIDIA and Simon Fraser University's Three-Step Approach:

• Step 1: Utilize Motion Capture Data: Start with a small dataset (14 minutes) of motion capture data from real humans.
• Step 2: Generate Random Levels: Create new, randomly generated game levels.
• Step 3: Physics-Based Motion Generation: Use a physics engine to generate new motions based on the existing data and the new levels. This corrects kinematic motions (AI-dreamed motions) that may involve floating or foot sliding.

3. The Iterative Training Process:

• The AI generates new kinematic motions, which are then corrected by the physics engine to ensure physical plausibility.
• These corrected motions are added to the initial dataset, enriching it.
• The cycle repeats, continuously improving the AI's abilities.

4. Path Generation and Character Actions:

• The AI generates paths within the levels that the character must follow.
• These paths involve actions like climbing and jumping.

5. Results and Learning:

• Initial training cycles yield imperfect results.
• After three iterations of dataset enrichment with physics-based correction, the AI demonstrates significant improvement.
• The AI learns to combine multiple motions, such as jumping, grabbing ledges, and climbing.
• The AI can perform complex actions and navigate challenging environments.

6. Testing on New Environments:

• The AI is tested on environments it has never seen before to assess its true intelligence.
• The "green character" represents the AI's initial, uncorrected motion, while the "blue character" represents the physics-corrected motion.
• The AI demonstrates the ability to adapt to new challenges, including climbing monuments and performing advanced maneuvers.
• Example: The AI hops forward on one leg after a jump, showcasing natural and adaptive movement.

7. Data Enrichment and Training Requirements:

• Each clip in the original motion capture dataset is converted into 50 different terrain variations.
• The training process can be performed on a single, albeit high-end (A6000), GPU.
• Training can take up to a month.

8. Limitations:

• The motion generation process is slow, requiring approximately 25 seconds of GPU processing time to create 1 second of character movement.

9. Conclusion:

• The research demonstrates the possibility of teaching an AI to navigate complex environments using a combination of motion capture data, physics-based simulation, and iterative dataset enrichment.
• The technology has the potential to be implemented in games and virtual worlds.
• The presenter expresses surprise that this work isn't more widely discussed.