University of California, Irvine researchers have built GLIDE, a system that turns repeated robot failures into executable motion guardrails. In three manipulation tasks that human operators could complete in no more than one of ten unassisted attempts, the refined guardrails raised successful demonstration collection to seven, nine and nine attempts out of ten.
Robot policies often learn by copying demonstrations collected through teleoperation. That approach assumes a person can reliably steer the machine through the task. The assumption failed here: operators had to carry a plate of round tomatoes without tilting it, hand a marker between two grippers and leave it standing, or coordinate a bottle and a dexterous hand to pour into a glass. Small timing or alignment errors spoiled an attempt before it could become useful training data.
Filtering commands before failure
GLIDE places a software filter between the operator and the robot. A coding agent starts from the task description and the existing teleoperation interface, then writes rules that can limit arm pose, wrist rotation, gripper state or finger motion. The filter uses the robot's own joint and motion state to decide when a risky command should be changed. After each round, GLIDE examines recorded video, proposed commands, executed commands and the outcome, then revises those rules.
The revisions created task phases that were absent from the initial rules. During plate transport, the two arms moved around a shared midpoint so one hand could not lift faster than the other. Marker placement gained a controlled release and retreat phase. Wine serving gained bottle-to-glass alignment and phase-aware tilt assistance. The researchers developed GLIDE's guardrails separately from a set written by domain experts, so the comparison did not begin with the expert rules embedded in the system.
What the robot trials showed
Each collection condition was tested in ten trials per task. Direct VR teleoperation succeeded in zero plate transfers, one marker task and zero pours. Expert-written guardrails lifted those results to two, six and six successes. The final GLIDE guardrails reached seven, nine and nine. The improvement was not steady in every refinement round: the wine task stayed at zero through the first two versions before the next revision produced nine successes. That trajectory shows both the value of feedback and the risk of treating an early generated rule set as finished.
The team also trained robot policies on the collected episodes and kept the same guardrails active during execution. With a mixture of successful and imperfect demonstrations, guarded policies completed seven plate transfers, six marker tasks and six pours in ten trials. Removing the runtime filter from those same trained policies reduced the results to zero, four and four. Guardrails alone were not enough either: policies trained only on unassisted demonstrations remained much weaker, even when the filter was added later.
Limits of the result
The tests cover three laboratory tasks, and every reported success rate is based on ten evaluation trials. At runtime, the guardrails read proprioception, meaning the robot's own position and motion signals, but do not see slipping objects, liquid movement or contact through vision and touch. Longer tasks, mobile robots and more varied hardware remain untested. GLIDE therefore shows a way to rescue demonstration collection when human control is unreliable; it does not yet establish that generated guardrails will generalize across robots or unfamiliar environments.
Sources
- Song, Mittal and Jain, “Learning Beyond What Humans Can Demonstrate,” arXiv:2609.24996 (submitted September 21, 2026)
- Full paper, including methods, Tables 1–2 and limitations
- GLIDE project page with task videos and guardrail revisions
- Lead image: Figure 1 from the paper by Yuchen Song, Aditya Mittal and Unnat Jain, reused under the paper's CC BY 4.0 license.