calvin

Why we included this project

Training robot-manipulation policies from natural-language instructions is a demanding problem, and CALVIN provides a simulated benchmark built for it. The agent has to execute a chain of commands in sequence, such as opening a drawer, picking up a blue block, and pushing it inside, which is far harder than the single short tasks found in most vision-and-language datasets. Teams working on language-conditioned control or embodied AI get a reproducible evaluation protocol with flexible sensor suites and multiple action spaces. The repository also includes baseline training code built on PyTorch Lightning and Hydra, so you can compare your own approach against established results instead of building an environment from scratch. That combination of a structured challenge and ready-made baselines makes it a practical place to validate long-horizon manipulation agents.