I3L combines human corrections and retraining for whole-body robot tasks

The I3L team introduced a system for correcting robot policies during long-horizon, whole-body tasks. Operators can take over through its JoyLo+ interface, and their corrections feed into retraining. The researchers report evaluating I3L across three robot embodiments and eight simulated and real-world tasks. Under a four-round HG-DAgger protocol, they report at least 90% autonomous success on every task and an average 68.3-percentage-point improvement over base policies.