Definition
Zero-shot transfer evaluates whether a policy can succeed on tasks or environments not seen during training. For robot foundation models, this means executing a new language instruction (e.g., "open the drawer") without any demonstrations of that specific task. Large VLA models like RT-2 and OpenVLA demonstrate zero-shot generalization to novel objects and instructions by leveraging visual and semantic knowledge from internet-scale pretraining. In sim-to-real, zero-shot transfer means deploying a simulation-trained policy directly in the real world without fine-tuning. Domain randomization and system identification are key enablers.
Why It Matters for Robot Teams
Understanding zero-shot transfer is essential for teams building real-world robot systems. Whether you are collecting demonstration data, training policies in simulation, or deploying in production, this concept directly affects your workflow and system design.