Meta AI unveiled a new algorithm that enables robots to learn and replicate human actions by watching YouTube videos. In a recent paper entitled “Affordances from Human Videos as a Versatile Representation for Robotics,” the authors explore how videos of human interactions can be leveraged to train robots to perform complex tasks.
This research aims to bridge the gap between static datasets and real-world robot applications. While previous models have shown success on static datasets, applying these models directly to robots has remained a challenge. The researchers propose training a visual affordance model using internet videos of human behavior could be a solution. This model estimates where and how a human is likely to interact in a scene, providing valuable information for robots.
The concept of “affordances” is central to this approach. Affordances refer to the potential actions or interactions an object or environment offers. By understanding affordances through human videos, the robot gains a versatile representation that enables it to perform various complex tasks. The researchers integrate their affordance model with four different robot learning paradigms: offline imitation learning, exploration, goal-conditioned learning, and action parameterization for reinforcement learning.
To extract affordances, the researchers utilize large-scale human video datasets like Ego4D and Epic Kitchens. They employ off-the-shelf hand-object interaction detectors to identify the contact region and track the wrist’s trajectory after contact. However, an important challenge arises when the human is still present in the scene, causing a distribution shift. To address this, the researchers use available camera information to project the contact points and post-contact trajectory to a human-agnostic frame, which serves as input to their model.
Previously, robots were capable of mimicking actions, but their abilities were limited to replicating specific environments. With the latest algorithm, researchers have made significant progress in “generalizing” robot actions. Robots can now apply their acquired knowledge in new and unfamiliar environments. This achievement aligns with the vision of achieving Artificial General Intelligence (AGI) as advocated by AI researcher Jan LeCun.
Meta AI is committed to advancing the field of computer vision and is planning to share its project’s code and dataset. This will enable other researchers and developers to further explore and build upon this technology. With increased access to the code and dataset, the development of self-learning robots capable of acquiring new skills from YouTube videos will continue to progress.
By leveraging the vast amount of online instructional videos, robots can become more versatile and adaptable in various environments.
Read more about AI:
Read More: mpost.io
Bitcoin 
Ethereum 
Tether 
XRP 
BNB 
Solana 
USDC 
Dogecoin 
Cardano 
TRON 
Lido Staked Ether 
Wrapped Bitcoin 
Sui 
Chainlink 
Avalanche 
Stellar 
LEO Token 
Toncoin 
Hedera 
Shiba Inu 
Wrapped stETH 
USDS 
Bitcoin Cash 
Litecoin 
Polkadot 
Hyperliquid 
Monero 
Binance Bridged USDT (BNB Smart Chain) 
WETH 
Bitget Token 
Ethena USDe 
Pi Network 
WhiteBIT Coin 
Wrapped eETH 
Coinbase Wrapped BTC 
Pepe 
Aptos 
Dai 
Uniswap 
OKB 
Bittensor 
NEAR Protocol 
Ondo 
Official Trump 
Gate 
Internet Computer 
sUSDS 
Aave 
BlackRock USD Institutional Digital Liquidity Fund 
Ethereum Classic