REDIRECT: A 1% Fix for Bad Robot Habits

Yu Zhang, Jiazhuo Li, Yancong Wei, Kangkang Dong, Xiaojun Zhu, Houde Liu · N/A · 2026

Summary

Abstract Summary

Robots can acquire bad habits from a few defective moments in otherwise useful teleoperation. In mixed-quality robot data, normal and problematic demonstrations share most task behavior and differ only at a local action continuation. Full retraining is costly, while fine-tuning on clean data alone offers limited recovery under a small update budget. We ask whether the local difference can instead be fixed with only 1% of the full-training sample-backward budget. Using only episode-level retained/problematic labels, REDIRECT localizes the branch, assigns a coherent retained continuation to the original observations around it, and anchors shared behavior, without frame-level annotations or additional interaction. Across three randomized ManiSkill tasks and three seeds, REDIRECT raises the mean success rate from 68.2% to 90.3%, recovering 88.8% of the clean-retraining gap versus 22.1% for matched-compute fine-tuning. On a PiPER arm, it recovers 86.7-92.9% of the clean-only gap across cup insertion and towel folding. Local robot habits can therefore be repaired by spending the update budget on the behavioral difference rather than relearning shared behavior.

Key Points

Robots can acquire bad habits from a few defective moments in otherwise useful teleoperation
In mixed-quality robot data, normal and problematic demonstrations share most task behavior and differ only at a local action continuation.
Full retraining is costly, while fine-tuning on clean data alone offers limited recovery under a small update budget.
We ask whether the local difference can instead be fixed with only 1% of the full-training sample-backward budget.
Using only episode-level retained/problematic labels, REDIRECT localizes the branch, assigns a coherent retained continuation to the original observations around it, and anchors shared behavior, without frame-level annotations or additional interaction.

Redirect: A One-Percent Behavioral Fix

Mixed-quality teleoperation structure. Let $\mathcal{D}_{r}=\{\tau_{i}^{r}\}$ contain ordinary demonstrations and $\mathcal{D}_{d}=\{\tau_{i}^{d}\}$ locally flawed demonstrations used to train $\pi_{\theta_{0}}$. Both execute the same task and share most behavior.

A trajectory in $\mathcal{D}_{d}$ follows ordinary behavior until a local decision departs from a continuation in $\mathcal{D}_{r}$; downstream observations can then diverge. In this setting, the local departure naturally induces a behavioral branch.

Redirect: A One-Percent Behavioral Fix — REDIRECT: A 1% Fix for Bad Robot Habits — Fig. 2: Randomized ManiSkill tasks with paired clean and problematic local behaviors. Correct (cyan) and erroneous (purple) operation points are marked in each scene; the lower-left and lower-right insets show matched clean and problematic intermediate states. PushCube arrows connect the current gripper center to the aligned or offset endpoint, while StackCube uses a side view to expose placement alignment.

Experimental Setup

We evaluate repair performance, computational cost, and behavioral preservation across three simulation tasks and two real-robot tasks using ACT and $\pi_{0.5}$. Each setting starts from a policy $\theta_{0}$ trained on 100 clean and 30 designated problematic demonstrations. Repair receives episode membership without temporal annotations. Simulation.

We use ManiSkill PickCube, PushCube, and StackCube [ 28 ] , with local grasp, push, and placement offsets of 50/60/50 mm, respectively.

Experimental Setup — REDIRECT: A 1% Fix for Bad Robot Habits — Fig. 3: Automatic branch localization and repair return. At matched sample-backward budgets, Redirect converts a 1%-scale update into substantially more clean-gap recovery than ordinary fine-tuning across simulation and real-robot settings; open circles denote matched clean FT, filled squares denote Redirect , and the triangle denotes four-times-SB $\pi_{0.5}$ fine-tuning. Panel (b) shows the empirical cumulative distribution of absolute localization error over 90 designated simulation episodes; the shaded band marks the $\pm 5$-frame tolerance, and synthetic branch timestamps are used only for audit.

Results

At the same 4,000-SB budget, Redirect raises mean success from 68.2% to 90.3%, recovering 88.8% of the clean-only gap; FT-CM recovers 22.1% (Table I ; Fig. 3 (a)). Redirect performs best among post-training methods on all three tasks, with a 5.0% regression rate (Table II ). The improvement also holds when different episodes contain different defect types.

On heterogeneous StackCube, Redirect reaches 81.0% success and recovers 68.6% of the clean-only gap without defect-family labels or retuning (Table V ).

Results — REDIRECT: A 1% Fix for Bad Robot Habits — Fig. 4: Branch-local policy diagnostics on problematic real trajectories. (a) ACT branch-relative original-label loss shift from $\theta_{0}$ (0%) toward clean-only (100%); shading marks BRANCH observations. (b) $\pi_{0.5}$ problematic-pattern assignment and clean-directed action progress. Error bars in (a) and brackets in (b) are 95% episode-bootstrap intervals.

Sources and Demonstrations

Figures are reproduced from Zhang et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.