NVIDIA’s AI Finally Solved Walking In Games

2025-12-21 · Source: Two Minute Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Gaming & Interactive Media, Robotics & Autonomous Systems · Depth: Intermediate, medium

Summary

NVIDIA researchers have developed an AI system that enables physically simulated characters to perform realistic walking and crowd behaviors in virtual environments, overcoming common "moonwalking" or "sliding feet" bugs seen in traditional game animations. Unlike conventional methods where NPCs are floating capsules with overlaid animations, this system uses agents with approximately 20 motor-driven joints that are physically simulated, allowing for natural adaptation to diverse terrains like stairs and slopes without pre-canned animations. The system comprises "Trace," a diffusion model that generates predictive paths, and "Pacer," a muscle-like component that controls joint movements to follow these paths while reacting to real-time physics. This approach, trained using Adversarial Reinforcement Learning with thousands of humanoids in parallel, produces organic crowd dynamics and realistic pedestrian behavior, with the source code freely available.

Key takeaway

For AI Scientists developing simulation environments or character animation systems, this NVIDIA research offers a robust alternative to traditional animation pipelines. You should explore integrating physically grounded AI agents, particularly for scenarios requiring realistic crowd behavior, terrain adaptation, or high-fidelity training data for autonomous systems. This approach can significantly enhance the realism and utility of your simulations, moving beyond static, rule-based character movements.

Key insights

NVIDIA's AI system uses physically simulated agents and diffusion models for realistic, adaptive character movement.

Principles

• Physical simulation enhances realism over canned animations.
• Generative pathfinding improves crowd behavior.
• Adversarial RL enables natural movement learning.

Method

The system combines a diffusion model ("Trace") for predictive path generation with a real-time physics-based joint controller ("Pacer"). These components interact, allowing the "muscle" to request new paths from the "brain" when encountering obstacles.

In practice

• Generate realistic pedestrian data for autonomous vehicle training.
• Create dynamic, adaptive NPC behaviors in games.
• Simulate diverse body types with natural movement.

Topics

• Physically Simulated Agents
• Adversarial Reinforcement Learning
• Diffusion Models
• Character Animation
• Autonomous Vehicle Simulation

Best for: AI Scientist, AI Engineer, Software Engineer, Research Scientist

Related on AIssential

• NVIDIA’s Insane AI Found The Math Of Reality — NVIDIA's PPISP technique improves NeRFs by mathematically correcting camera-induced distortions in input images.
• NVIDIA Research Advances Robotics From Simulation to the Real World — Simulation-to-real transfer is foundational for achieving generalizable, reliable embodied autonomy in robotics.
• Disney Characters Coming to Life at NVIDIA GTC 2026 — Disney uses physical AI, Newton, and Kamino to create expressive, autonomous robotic characters.
• Verkada Accelerates Physical AI with NVIDIA — NVIDIA's investment and technical collaboration with Verkada significantly enhance physical AI capabilities for real-world operational intelligence.
• Nvidia Launches Physical AI Models for Robots — Nvidia's new physical AI models and frameworks aim to democratize intelligent robot development and deployment.
• AI Generates Playable Interactive environments from an image in Real-Time - This Changes Everything — Mirage AI generates real-time, interactive game worlds from images, trained on human gameplay.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Two Minute Papers.