Dataconomy
  • News
    • Artificial Intelligence
    • Cybersecurity
    • DeFi & Blockchain
    • Finance
    • Gaming
    • Startups
    • Tech
  • Industry
  • Research
  • Resources
    • Articles
    • Guides
    • Case Studies
    • Glossary
    • Whitepapers
  • Newsletter
  • + More
    • Conversations
    • Events
    • About
      • About
      • Contact
      • Imprint
      • Legal & Privacy
      • Partner With Us
Subscribe
No Result
View All Result
  • AI
  • Tech
  • Cybersecurity
  • Finance
  • DeFi & Blockchain
  • Startups
  • Gaming
Dataconomy
  • News
    • Artificial Intelligence
    • Cybersecurity
    • DeFi & Blockchain
    • Finance
    • Gaming
    • Startups
    • Tech
  • Industry
  • Research
  • Resources
    • Articles
    • Guides
    • Case Studies
    • Glossary
    • Whitepapers
  • Newsletter
  • + More
    • Conversations
    • Events
    • About
      • About
      • Contact
      • Imprint
      • Legal & Privacy
      • Partner With Us
Subscribe
No Result
View All Result
Dataconomy
No Result
View All Result

AI is learning to drive like a human—by watching you panic

Proxy Value Propagation (PVP) is a new method that turns AI training into something far more human

byKerem Gülen
February 6, 2025
in Research

Self-driving cars are supposed to be the future. AI is supposed to take the wheel, navigate flawlessly, and eliminate human error. Yet here we are, still gripping our steering wheels while AI stumbles through simulations, making mistakes that range from hilariously bad to downright dangerous.

Why? Because AI learns through trial and error—the digital equivalent of throwing darts in the dark until it finally hits the bullseye. That’s fine when the stakes are low, like playing chess or optimizing ads. But when it comes to real-world applications—where a mistake means plowing into a pedestrian—this approach falls apart.

According to a study conducted by Zhenghao Peng, Wenjie Mo, Chenda Duan, and Bolei Zhou from the University of California, Los Angeles (UCLA), along with Quanyi Li from the University of Edinburgh, AI training can be dramatically improved using Proxy Value Propagation (PVP). Their research, titled Learning from Active Human Involvement through Proxy Value Propagation, challenges traditional reinforcement learning by proving that active human intervention leads to faster, safer, and more efficient AI training.

Stay Ahead of the Curve!

Don't miss out on the latest insights, trends, and analysis in the world of data, technology, and startups. Subscribe to our newsletter and get exclusive content delivered straight to your inbox.

Traditional Reinforcement Learning (RL), the standard way AI learns to make decisions, is painfully slow. It requires millions of attempts before an AI figures out what works. Worse, it assumes AI can understand human intent just by following a reward system—when in reality, reward systems often lead to bizarre, unintended behaviors. Think of an AI trained to win a race that figures out it can just drive in circles at the start line to rack up “distance traveled” points without ever finishing the course.

Clearly, AI needs a better teacher. And that teacher? You.

Let humans intervene in real time

Proxy Value Propagation (PVP) is a new method that turns AI training into something far more human. Instead of letting AI blunder through its mistakes for months, PVP lets humans step in, intervene, and show AI what to do in real time.

  • Imagine AI is learning to drive in a simulation, say, Grand Theft Auto V (GTA V).
  • The AI makes a terrible decision—say, running a red light straight into traffic.
  • Instead of watching the chaos unfold, a human takes control at that moment and corrects the AI’s action.
  • The system then labels the human’s decision as a “good” move and the AI’s previous mistake as a “bad” move.
  • Using a technique called value propagation, AI spreads this correction across similar situations, learning to avoid bad decisions without needing millions of attempts.

The result is surprising. AI learns much faster, with fewer mistakes, and—most importantly—it actually aligns with human expectations instead of blindly chasing reward points.


AI struggles with strategy: Study shows LLMs reveal too much in social deduction games


The numbers don’t lie: PVP works

The team behind PVP put it to the test in GTA V, CARLA (a driving simulator), and MiniGrid (a virtual maze navigation task). The results were stunning:

  • AI trained with PVP learned 10 times faster than traditional methods.
  • It required only 1,200 human interventions—compared to the 300,000 attempts AI typically needs in RL.
  • The success rate of PVP-trained AI in reaching destinations safely was 85%, compared to just 20-50% for previous methods.
  • AI made 75% fewer critical mistakes when trained with PVP versus traditional reinforcement learning.

In other words, AI actually started driving like a human—not just a robot programmed to maximize abstract rewards.

A win for AI—and for humans

PVP isn’t just better for AI. It also makes life easier for the people training it. Traditional AI training requires constant human oversight, hours of feedback, and a whole lot of patience. With PVP, AI needed 50% less human effort to train. Testers rated PVP-trained AI 4.8 out of 5 for accuracy, compared to just 3.0 for older methods. AI that followed PVP training caused significantly less stress for human trainers—because it didn’t constantly require corrections. For a technology that’s supposed to make our lives easier, that’s a huge step forward.

From GTA to the streets

PVP has already proven itself in virtual driving tests. The real question is: can it work in real-world applications?

The potential is massive. Instead of relying solely on pre-programmed rules, AI could learn directly from human intervention—making it safer, faster. AI-powered robots in warehouses, hospitals, or even homes could be trained in real time instead of through trial-and-error. Human doctors could intervene during AI-assisted surgeries or diagnoses, directly teaching the system what’s right or wrong.

Sometimes, the goal is just to make AI human enough—to act in ways we expect, to align with our values, and to avoid mistakes that put us at risk.


Featured image credit: Kerem Gülen/Midjourney

Tags: AIFeatured

Related Posts

Researchers make breakthrough in semiconductor technology set to supercharge 6G

Researchers make breakthrough in semiconductor technology set to supercharge 6G

October 3, 2025
Light-powered chip makes AI computation 100 times more efficient

Light-powered chip makes AI computation 100 times more efficient

October 3, 2025
Researchers combine OLEDs and metasurfaces to advance holographic displays

Researchers combine OLEDs and metasurfaces to advance holographic displays

October 1, 2025
Diraq, Imec show 99% fidelity in silicon qubit production

Diraq, Imec show 99% fidelity in silicon qubit production

October 1, 2025
CDU study: AI threatens human dignity globally

CDU study: AI threatens human dignity globally

October 1, 2025
HBR: Dysfunctional idea marketplaces stall innovation

HBR: Dysfunctional idea marketplaces stall innovation

October 1, 2025

LATEST NEWS

ChatGPT reportedly reduces reliance on Reddit as a data source

Perplexity makes Comet AI browser free, launches background assistant and Chess.com partnership

Light-powered chip makes AI computation 100 times more efficient

Free and effective anti-robocall tools are now available

Choosing the right Web3 server: OVHcloud options for startups to enterprises

Z.AI GLM-4.6 boosts context window to 200K tokens

Dataconomy

COPYRIGHT © DATACONOMY MEDIA GMBH, ALL RIGHTS RESERVED.

  • About
  • Imprint
  • Contact
  • Legal & Privacy

Follow Us

  • News
    • Artificial Intelligence
    • Cybersecurity
    • DeFi & Blockchain
    • Finance
    • Gaming
    • Startups
    • Tech
  • Industry
  • Research
  • Resources
    • Articles
    • Guides
    • Case Studies
    • Glossary
    • Whitepapers
  • Newsletter
  • + More
    • Conversations
    • Events
    • About
      • About
      • Contact
      • Imprint
      • Legal & Privacy
      • Partner With Us
No Result
View All Result
Subscribe

This website uses cookies. By continuing to use this website you are giving consent to cookies being used. Visit our Privacy Policy.