What will be the next big breakthrough in decision making / robotic learning / reinforcement learning in the next 5 years? (think e.g. AlexNet, ResNets, Transformers, Diffusion)

What is something we do right now in decision making / robotic learning / reinforcement learning that we WILL NOT be doing 5 years ? (think e.g. handcrafted features,)

RL algorithms for autonomously influencing humans using social media, via auto generated content

The benchmark everyone will be working on will be on assistive human computer interfaces

trying to learn on a real robot from scratch in the real world

Trying to do reset-free RL

Learning exploration as an accumulating "experience"

model-based rl (at least on high-level decision making)

Working with fixed rewards

Applications in hard-to-model domains that are not embodied

Robots learning with self play: think two physical Robots trying to be cooperative/adversarial (their choice) in maximizing their skill acquisition speed. We’ll see cooperatives, deception, murder, and everything in between!

We will really not be training from scratch or designing reward functions. Everything from natural language: describe desired behavior and condition on it!

State entropy based exploration

Training large models without backpropagation

Internet scale exploration policies to aid foundation models data acquisition

RL + traditional robotics (we will codevelop hardware specifically for RL)

Use big generative models pretrained on text or images from non-interactive internet sources

some way to make RL more like supervised learning

thinking carefully about the composition or deists of robotics datasets (i.e we will have methods that can learn from very messy/heterogenous data (different controllers, cameras, robots))

A structural separation of decision making and control

Using rewards as the source of supervision

Pre-trained visual representations for RL

Distinguish the learning algorithm between online/offline learning

Submitting to conferences

RL algorithms for autonomously influencing humans using social media, via auto generated content

large scale offline pretraining

The benchmark everyone will be working on will be on assistive human computer interfaces

Simple methods massively scaled up

Learning exploration as an accumulating "experience"

probabilistic programming

Applications in hard-to-model domains that are not embodied