Nadun and Juny, second-year PhD students at Georgia Tech, explain SAIL, their co-authored approach to faster imitation-learning policy execution, with RoboPapers host Michael.
The researchers say imitation-learning policies often run at the pace of cautious human teleoperation, while industrial robots have long operated at high speed. SAIL asks whether visuomotor policies can execute faster than their demonstrations without losing reliability.
Running action chunks faster can leave the robot waiting for inference, causing abrupt stops and jerky motion. Asynchronous execution lets the robot act while the next chunk is generated, but policy predictions can diverge and make the robot oscillate. Longer chunks reduce waiting but can be less accurate and less responsive to changes.
SAIL uses error-adaptive guidance to make new action chunks agree with the previous trajectory, reducing disruptive jumps between predictions. The policy is trained to condition on future actions, and guidance can be reduced when tracking error calls for a more reactive response. Motion-complexity labels let the robot move faster in free space and slow down near tasks such as grasping.
An action scheduler accounts for inference delay so the robot can keep enough actions available to avoid pauses. The team also uses higher controller gains, trains on poses the robot actually reached, and applies feed-forward control. Together, these choices produced nearly three times the baseline throughput, measured as successful attempts per unit time.
Adaptive slowdown can reduce throughput on complex tasks, but may prevent costly failures such as a bad peg insertion. The demonstrations should be successful and relatively smooth, though they can vary in how the task is performed. The team notes torque limits and safety as challenges when moving faster, and plans to explore industrial robots and real-time safety controllers.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Seonghoon, a PhD student at KAIST and research intern at NVIDIA, explains DreamZero, a world action model for zero-shot robot policies, in a Robo Papers conversation with host Michael.
Jason Ma, a robotics researcher at the University of Pennsylvania who worked on the paper at Google DeepMind, joins RoboPapers hosts Michael and Chris to discuss using vision-language models to assess robot task progress and data quality.
