Learning Beyond What Humans Can Demonstrate

TL;DR

GLIDE enables robots to learn manipulation tasks when human demonstrations are impractical by automatically generating dynamic guardrails that filter unsafe actions during teleoperation and policy execution. It uses task descriptions and teleoperation code to infer failure modes and iteratively refine constraints, significantly boosting successful data collection and policy performance. The framework achieves 70–90% success in data collection and 60–70% task success across three challenging manipulation tasks.

Why It Matters

Many complex robotic tasks, such as precise handovers or stable object transfers, are difficult for humans to demonstrate reliably. GLIDE overcomes this bottleneck by enabling efficient data collection and policy learning without requiring perfect human demonstrations, making advanced robot learning feasible in real-world scenarios.

Key Findings

• GLIDE automatically infers task-specific failure modes from task descriptions and teleoperation code to generate dynamic guardrails.
• Guardrails improve data collection success from 0–10% to 70–90% across three tasks, outperforming both naive teleoperation and hardcoded expert rules.
• Policies trained on mixed data with GLIDE's guardrails achieve 60–70% success rates on challenging tasks like wine serving and marker handover.