RL@B
Apply

Research

Applied RL: simulation, a real robot arm, and RL for language models. Pods form around real problems.

Focus areas

01

Robot learning

Teaching a low-cost robot arm from human demonstrations and from simulation.

Imitation learningSimulation to realOffline RL
02

Simulation and world models

Fast simulated environments, and learned models of the world that an agent can plan in.

Physics simulationWorld modelsPlanning
03

RL for language models

Training language models on tasks with checkable answers, with scoring that can't be gamed.

Verifiable rewardsReward modelsRobust scoring
04

Evaluation and benchmarks

Benchmarks for long tasks, rubrics graders agree on, and measuring that agreement.

Long tasksRubricsAgreement

Research process

  1. 01

    Reading group

    Weekly. One paper, one presenter.

  2. 02

    Research pods

    Formed around a problem or a release.

  3. 03

    Capstones to papers

    DeCal projects continue as workshop papers.

  4. 04

    Published openly

    Environments and course materials are released publicly.

DeCal

Applied RL

Spring 2027 · 2 units · P/NP

From the fundamentals to trained agents in 15 weeks. Every module ends in a working project.

Open workshops

  • Your first RL environment
  • Train a language model with RL in 90 minutes
  • Tasks, rubrics and answer keys (no programming)
  • Robot arm from scratch
  • Setting up a physics simulator

Syllabus · 15 weeks

  1. Wk 1 to 2Foundations: decision processes and value functions
  2. Wk 3 to 5Deep RL: value-based and policy-gradient methods
  3. Wk 6 to 7Continuous control and reward design
  4. Wk 8 to 9RL for language models
  5. Wk 10 to 11Robot learning, offline RL and world models
  6. Wk 12How frontier labs do RL (guest lecture)
  7. Wk 13 to 15Project sprint and demo day

Research pods

Build-track members join a pod in their first semester. Labs and companies can propose a problem.