Anonymous Author(s)

Episodes A and B play real teleoperated episodes while the curve shows the dense per-frame predicted signed progress magnitude \(\hat{v}_{t}\) — positive for forward task progress and negative for regression. Click on the curve to scrub through the video.

Data

Our policy-training data is drawn from a single corpus of around 125 hours of successful, unannotated human-teleoperated T-shirt-folding demonstrations. On this task, episode length is a coarse proxy for execution efficiency: longer episodes tend to contain more hesitations, retries, and recoveries. To evaluate robustness as progressively more inefficient behavior is admitted into training, we define three nested, length-filtered tiers: \(\mathcal{D}_{1}\) (≤ 60s): 1,975 episodes (29.1 hours), \(\mathcal{D}_{2}\) (≤ 90s): 3,546 episodes (61.5 hours), and \(\mathcal{D}_{3}\) (≤ 120s): 5,718 episodes (124.7 hours). Policies are trained on the same underlying demonstrations, either with uniform weighting (vanilla BC) or WARP-based progress reweighting. A single WARP-RM model is used across all tiers: it is trained once on a fixed reference subset \(\mathcal{D}_{\mathrm{RM}}\) — the shortest demonstrations (≤ 59s, 1,807 episodes) — providing a clean reference signal for the canonical execution pace (\(\hat{v} = 1\)). For baseline comparisons, SARM requires human annotations, so an annotated supplement \(\mathcal{D}_{A}\) (929 expert demonstrations, 15.8 hours) is added, forming the augmented datasets \(\mathcal{D}_{4} = \mathcal{D}_{1} \cup \mathcal{D}_{A}\) and \(\mathcal{D}_{5} = \mathcal{D}_{2} \cup \mathcal{D}_{A}\); WARP-RM and all other baselines treat \(\mathcal{D}_{A}\) as unannotated.

Real-World Policy Rollouts

Across 380 real-world trials of T-shirt folding from a crumpled start, WARP-BC consistently completes more folds and completes them faster than vanilla BC and other baselines trained on the same demonstration corpus. Each video below shows all 20 evaluation trials for a tier, played simultaneously in a 4×5 grid. Use the selector to switch between models trained on each of the three demonstration tiers from the paper.

Videos played at 1× speed.

Method

Prior progress models are trained to predict how far along a trajectory is — the fraction of the demonstration that has elapsed so far. This is a noisy learning target, because the same elapsed fraction in two different demonstrations can correspond to very different amounts of actual task progress. WARP-RM instead learns a relative signal — how fast, and in which direction, the task is advancing — using a self-supervised target obtained by replaying successful demonstrations at non-uniform velocities.

2. Predict per-frame velocity from images

A frozen DINOv3 ViT-B/16 + 12-layer bidirectional transformer head outputs a per-frame categorical distribution over cumulative progress. Its temporal derivative gives the velocity \(\hat{v}_{t}\):

\(\hat{v}\) ≈ 1 → expert pace, \(\hat{v}\) ≈ 0 → stagnating, \(\hat{v}\) < 0 → regressing

Quantitative Results

Cross-tier results on T-shirt folding

All policies are evaluated on 20 trials of T-shirt folding from a crumpled start with a 240s timeout. Mean time-to-completion (TTC) is reported over successful trials only. As the training pool admits more suboptimal demonstrations, vanilla BC degrades sharply while WARP-BC stays robust.

Matched baseline comparisons

Because SARM requires human-annotated subtask boundaries, all methods are evaluated on the augmented corpora \(\mathcal{D}_{4} = \mathcal{D}_{1} \cup \mathcal{D}_{A}\) and \(\mathcal{D}_{5} = \mathcal{D}_{2} \cup \mathcal{D}_{A}\), where \(\mathcal{D}_{A}\) is the annotated supplement (treated as unannotated by every method except SARM). WARP-BC sustains the highest throughput on both tiers and ties or leads on success — without the human labels SARM needs. SARM and SCIZOR collapse on the noisier \(\mathcal{D}_{5}\), while DemInf stays robust on success but at lower throughput.

Ablations

All ablations are run on dataset \(\mathcal{D}_{2}\). Kept-train-samples reports the fraction of action chunks that survive the weighting filter.