NVIDIA’s New AI’s Movements Are So Real It’s Uncanny

Why It Matters

ADD automates the reward‑design step in physics‑based animation, dramatically cutting development time and cost while delivering more realistic motion for games, film, and robotics.

Summary

The video spotlights a recent breakthrough in physics‑based character animation: the Adversarial Differential Discriminator (ADD). Building on the 2018 DeepMimic framework, which turned motion imitation into a video‑game‑style reward‑maximization problem, ADD replaces dozens of hand‑crafted score counters with a single learned AI judge that evaluates how closely a synthetic motion matches a human reference.

DeepMimic’s strength lay in its ability to reproduce captured motions, but it demanded painstaking manual tuning of rewards for joint angles, foot placement, balance, and more. ADD eliminates that bottleneck by training an adversarial discriminator to automatically infer what a “perfect” performance looks like, then feeding its feedback to the controller. In side‑by‑side tests the new method matches or exceeds DeepMimic on challenging tasks such as parkour jumps, climbing, and robot locomotion, while retaining the original’s flexibility across different body morphologies.

The presenter highlights vivid examples: DeepMimic collapses on a high‑energy jump that ADD executes fluidly, and an ablation study shows that removing any of ADD’s components degrades performance. He also notes a limitation—occasionally the AI judge falters on flashier tricks, likening it to a dance judge who freezes on a backflip. Nonetheless, the overall results demonstrate that a learned reward function can replace labor‑intensive hand‑tuning.

For the industry, this shift means animators, game studios, and robotics engineers can generate realistic, physically plausible motion with far less manual effort, accelerating development cycles and lowering costs. As AI judges become more adept, we can expect increasingly lifelike digital humans and more agile robotic controllers, reshaping entertainment, simulation, and autonomous systems.

Original Description

❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambda.ai/papers

Guide:

Rent one of their GPUs with over 16GB of VRAM

Open a terminal

Just get Ollama with this command - https://ollama.com/download/linux

Then run ollama run gpt-oss:120b - https://ollama.com/library/gpt-oss:120b

📝 The paper is available here:

https://add-moo.github.io/

📝 My paper on simulations that look almost like reality is available for free here:

https://rdcu.be/cWPfD

Or this is the orig. Nature Physics link with clickable citations:

https://www.nature.com/articles/s41567-022-01788-5

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:

Benji Rabhan, B Shang, Christian Ahlin, Gordon Child, John Le, Juan Benet, Kyle Davis, Loyal Alchemist, Lukas Biewald, Michael Tedder, Owen Skarpness, Richard Sundvall, Steef, Sven Pfiffner, Taras Bobrovytsky, Thomas Krcmar, Tybie Fitzhugh, Ueli Gallizzi

If you wish to appear here or pick up other perks, click here: https://www.patreon.com/TwoMinutePapers

My research: https://cg.tuwien.ac.at/~zsolnai/

X/Twitter: https://twitter.com/twominutepapers

Thumbnail design: Felícia Zsolnai-Fehér - http://felicia.hu

#nvidia