MinInter boosts robot imitation learning by minimizing interpolation in data augmentation

Imitation learning for robotics often struggles with the high cost of collecting high-quality demonstration data. Data augmentation methods combat this by recombining expert trajectories under different initial states, but they frequently introduce non-expert interpolated segments that degrade data quality. The new MinInter (Minimizing Interpolation) method, presented by Qingyang Wang and collaborators at IEEE CASE 2026, tackles this head-on. For each sampled initial configuration, MinInter selects the source demonstration that requires the least interpolation to form a complete trajectory. This explicit minimization of interpolation steps yields higher-quality synthetic demonstrations while remaining compatible with existing data generation pipelines.

Tested across 12 manipulation tasks with 26 variants from the MimicGen benchmark, MinInter consistently improved both data generation success rates and final policy success rates. The largest gains were observed in contact-rich, long-horizon, and high-variance settings — precisely the scenarios where interpolation artifacts cause the most harm. Notably, MinInter achieved higher policy success rates than the more complex SkillGen framework, proving that a simple focus on interpolation minimization can yield outsized benefits. This approach promises to accelerate robot skill acquisition by making data augmentation more effective without requiring more real-world demonstrations.