engblogs

Announcing the completion of open-sourcing MuJoCo, a physics simulator, with a roadmap for future collaboration and best-in-class capabilities.

The Berkeley Crossword Solver (BCS) combines neural question answering and probabilistic inference to achieve near-perfect performance on American-style crossword puzzles, outscoring human competitors and achieving 99.7% letter accuracy.

A blogpost featuring the journey of a software engineer on DeepMind's robotics team, from LEGO competitions to joining the company and experiencing a typical day at work.

A personal story about overcoming self-doubt to work at DeepMind's robotics lab.

Exploring the latest in DALL·E 2 research advancements.

Exploring how reinforcement learning agents learn bartering behavior and economic decision-making in a multi-agent environment.

The blogpost discusses how machine learning benchmarks in trading do not track what traders care about.

Exploring how populations of deep RL agents learn microeconomic behaviours, such as production, consumption, and trading of goods.

Applying a large-scale language model approach to build a single generalist agent capable of multi-modal, multi-task, multi-embodiment interactions.

Applying large-scale language modelling to build a multi-modal, multi-task, multi-embodiment generalist agent capable of playing games, captioning images, chatting, and performing physical tasks.

Active offline policy selection using offline data, special kernel, and Bayesian optimization for efficient RL policy evaluation

Using active offline policy selection (A-OPS) to improve the selection of RL policies for real-world applications by leveraging prerecorded datasets and limited interactions with the environment.