A$AP Robots

Happy Monday!

Last week, researchers from Carnegie Mellon University and NVIDIA unveiled ASAP (Aligning Simulation and Real-World Physics), a groundbreaking framework that's teaching humanoid robots to move with unprecedented human-like agility. But beyond the technical achievements lies a fascinating story about bridging the gap between virtual training and physical reality in robotics.

Here's what you need to know:

• ASAP enables humanoid robots to perform complex movements like Cristiano Ronaldo's signature jump celebration and LeBron James's "Silencer" move

• The system uses a two-stage approach: first training in simulation, then fine-tuning with real-world data

• This represents a major step toward robots that can move naturally in the real world, not just in simulations

• The implications extend far beyond sports moves, potentially revolutionizing fields from manufacturing to healthcare

The Two-Part Story

Part 1: The Virtual Training Ground

Think about learning to swim by first practicing on dry land. That's essentially what happens when robots learn movements in simulation - it's safe and efficient, but doesn't perfectly match real-world conditions. The ASAP team tackled this by first creating highly detailed simulations where robots could learn complex movements by watching human videos.

What makes this approach special is how it breaks down human movements into components that robots can understand and replicate. The system goes beyond copying surface-level movements; it learns the underlying principles of balance, momentum, and coordination that make these movements possible.

Part 2: Bridging the Reality Gap

Here's where things get really interesting. Instead of hoping the simulated training would work in the real world (spoiler alert: it usually doesn't), the ASAP team developed what they call a "delta action model." Think of it as a translator between the idealized world of simulation and the messy reality of physical hardware.

This model learns to compensate for all the little differences between simulation and reality - things like motor delays, friction variations, and the countless other small factors that make real-world movement challenging. It's like having an experienced coach who knows exactly how to adjust your textbook knowledge for practical application.

The Numbers That Matter

Here are some stats and metrics from this groundbreaking approach:

• Success Rate: Nearly 100% for many complex movements that were previously impossible

• Performance Improvement: Up to 52.7% reduction in motion tracking errors

• Real-World Validation: Successfully tested on the Unitree G1 humanoid robot across multiple challenging movements

Why This Changes Everything

The gap between simulation and reality has been one of the biggest roadblocks in robotics. While we've been able to create impressive simulations of robot movements for years, getting those same movements to work in the real world has been extraordinarily difficult.

ASAP's approach is significant because it’s incredibly scalable. Once the system is trained, it can help robots learn new movements quickly. It balances speed with practicality, as the framework works with existing robot hardware and doesn’t rely on added complexity. Additionally, this work is also adaptable to other types of robots and movements.

Looking Ahead

While ASAP represents a significant breakthrough, we're still in the early stages of this technology. The researchers note several challenges, including hardware limitations and the need for extensive real-world data collection. However, the path forward is clear: we're rapidly moving toward a future where robots can move as naturally as humans do.

Until next week, keep innovating.

1. OpenAI co-founder John Schulman leaves Anthropic after just five months (TC)

2. Nvidia boss Jensen Huang meets Donald Trump at White House (FT)

3. Lyft to bring Claude to more than 40 million riders (A)

4. Salesforce Cutting 1,000 Roles While Hiring for AI (BBG)

5. Rethinking the Scaling-Up of One-Stage Conditioned Human Animation Models (GH)

6. AI-powered map of the abdomen could help find cancer early on (MX)

7. Using a transformer to improve end of turn detection (LK)

8. Alibaba releases AI model it says surpasses DeepSeek (RT)

9. India now OpenAI's second largest market (RT)

10. Amazon's AI revamp of Alexa assistant nears unveiling (RT)

_{As a brief disclaimer I sometimes include links to products which may pay me a commission for their purchase. I only recommend products I personally use and believe in. The contents of this newsletter are my viewpoints and are not meant to be taken as investment advice in any capacity. Thanks for reading!}