Robots often learn by copying human demonstrations. But some tasks — those that need precise contact timing, fast balance, or very careful finger work — are hard or impossible for people to demonstrate by hand. This paper introduces GLIDE (Guardrails for Learning from Infeasible Demonstrations Efficiently), a system that writes short safety rules, or “guardrails,” that make remote teleoperation and robot policies safer and more usable for collecting training data.

Rather than requiring a perfect human demonstration, GLIDE takes a task description and the code used for teleoperation and looks for ways humans and naive teleoperation fail. In plain terms, it finds the actions that tend to cause failure and turns that knowledge into runtime rules. Those rules filter or change commands from a human operator or a learned policy during data collection and at deployment. Teleoperation means a human controls the robot remotely, often in virtual reality; the guardrails intervene when such control would likely cause failure.

At a high level GLIDE infers likely failure modes, writes checks based on the robot’s current state, and applies those checks to block or constrain risky motions. It also uses trajectory feedback — data about how attempts went — to refine the guardrails over time. The result is an iterative loop: try, see how it fails, and adjust the guardrails so future attempts avoid the same mistakes.

The authors tested GLIDE on three manipulation tasks: Tomato plate transfer, Marker handover and stand, and Wine serving. Data-collection success rose dramatically after guardrail refinement, from initial rates near 0–10% up to 70–90% across the tasks. When policies were trained on the guarded, mixed datasets and then run with guardrails, they achieved about 70%, 60%, and 60% success on the three tasks respectively. The paper reports that GLIDE improved over naive virtual-reality teleoperation and over hand-coded guardrails written by domain experts, and that it sometimes discovered useful guardrails that the experts had not thought of.