Justin Yu^1,3,*,†, Andrew Goldberg^1,*, Kavish Kondap^1,*, Karim El-Refai^1,*, Ethan Ransing¹, Qianzhong Chen², Mac Schwager², Fred Shentu³, Philipp Wu³, Ken Goldberg¹

¹University of California, Berkeley    ²Stanford University    ³XDOF

^*Equal contribution.    ^†Corresponding author: yujustin@berkeley.edu

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

Abstract ▾(click to expand)

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 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 the three demonstration tiers from the paper.

Every rollout below runs at 1× speed, fully autonomous.

Both videos are time-aligned: WARP-BC finishes folds while vanilla BC is often still hesitating or recovering.

Method

Prior progress models supervise on absolute normalized time, which is noisy: identical timesteps across demos can correspond to very different task states. WARP-RM instead learns a relative signal — how fast, and in which direction, the task is advancing — with supervision generated for free 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 → stalled · \(\hat{v}\) < 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. 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 D₄ = D₁ ∪ A and D₅ = D₂ ∪ A, where A is the annotated supplement (treated as unlabeled 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 D₅, while DemInf stays robust on success but at lower throughput.

Ablations

All ablations are run on dataset D₂. Kept-train-samples reports the fraction of action chunks that survive the weighting filter.