I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation

Policies trained offline from teleoperated demonstrations can degrade outside their training distribution. This gap is particularly acute in long-horizon whole-body mobile manipulation, where long horizons amplify compounding errors and coordinated arm, torso, and base motion makes exhaustive demonstration coverage impractical. We introduce I3L, an integrated hardware and learning system that couples targeted whole-body correction with iterative policy refinement to address this last-mile reliability gap. Its low-cost, open-source JoyLo+ interface uses policy-tracking leader arms and touch-triggered takeover to enable coordinated corrections across the arms, torso, and mobile base, while repeated deployment and retraining allow operators to target failures of the updated policy. We evaluate I3L across three robot embodiments and eight simulated and real-world tasks, including four long-horizon whole-body mobile manipulation tasks. Controlled simulation comparisons show that four-round HG-DAgger achieves the highest final success on both simulation tasks. Under this protocol, I3L improves autonomous success over the base policies by an average of 68.3 percentage points and achieves at least 90% success on every task, outperforming demonstration-only training by 35.6 percentage points on average under matched training-trajectory budgets. Across four rounds, task-averaged takeover duration per episode decreases by 72%. Together, these results show that I3L helps close the last-mile gap between offline training and reliable deployment and provides practical insights into collecting and using corrections for long-horizon whole-body mobile manipulation.

JoyLo+ builds on GELLO's kinematic leaders and JoyLo's actuated interface to let an operator step in during autonomous execution.

• Leaders track the policy. Motorized leaders mirror the policy's arm targets during autonomy and become gravity-compensated for hand guidance during a correction.
• Grasp to take over. Capacitive electrodes in the grips detect touch, so the operator takes control simply by holding the device.
• Whole-body takeover. On R1 Pro, one takeover hands over both arms, torso, base, and grippers; releasing returns control to the policy, which replans from the corrected state.

Starting from a demonstration dataset D₀, each round deploys the previous policy and collects a labeled rollout dataset through JoyLo+. Under HG-DAgger filtering, only the original demonstrations and the active intervention frames from all rounds are retained for training. Pre-intervention and on-policy frames are excluded to concentrate training on human demonstrations and corrective supervision.

Each round trains from the same starting weights rather than continuing from the previous round's policy: the same pretrained checkpoint for the π_0.5-based VLA, and the same initialization for diffusion policies. As earlier failures become less frequent, corrections naturally shift toward downstream stages such as transport and placement, producing an emergent failure-mode curriculum.

Eight tasks on three robots: two controlled BEHAVIOR simulation tasks for comparing collection and learning strategies, two contact-rich precision tasks on Franka, and four long-horizon whole-body mobile manipulation tasks on a real R1 Pro.

With the same trajectory budget, targeted corrections outperform additional demonstrations on every task and reach at least 90% autonomous success on all eight.

As the policy is retrained, operators step in less often, for less of each episode, and for less time, with all three measures declining every round.

As early-stage failures are fixed, the remaining corrections concentrate on each task's later stages, forming an emergent curriculum over failure modes.

Four-round HG-DAgger finishes best on both tasks, but more rounds alone do not guarantee a better policy: how corrections are learned from matters as much as collecting them.