wtf is srsa?

SRSA is a research project from NVIDIA Labs that helps robots learn new physical assembly tasks faster. Instead of training a robot from scratch for every new job, like fitting parts together or tightening components, it pulls a relevant existing skill from a library of pre-trained ones and adapts it to the current task. The system works in two main steps. First, it figures out which existing skill is the best match. It analyzes the geometry of the parts, how they move, and what actions the robot needs to take, then predicts which pre-trained skill is most likely to transfer successfully to the new task. For example, it might determine that a robot trained on assembly task 00783 is a great starting point for learning task 01036. Second, it takes that selected skill and fine-tunes it for the specific new task. The project includes a technique called "self-imitation learning," where the robot essentially learns from its own best attempts. Because it's starting from a strong foundation rather than a blank slate, the robot can learn effectively even with sparse feedback, meaning it doesn't need constant, detailed guidance on every move. This is primarily a tool for robotics researchers and engineers working in simulation environments, specifically NVIDIA's Isaac Lab and Isaac Sim. If you're building robotic assembly systems and want to explore how skill transfer can speed up training, this gives you the framework and pre-trained models to experiment with. It's a research paper implementation (accepted at ICLR 2025), so expect it to be more of a starting point for experimentation than a polished commercial product. Worth noting: the code is tightly coupled to NVIDIA's simulation stack and requires specific versions of Isaac Lab, Isaac Sim, and CUDA. It also includes pre-trained models and sample data to get you started. The self-imitation learning piece is a custom fork of an existing reinforcement learning library, so it's not meant to be a general-purpose replacement for that tool.