Robot demos no longer need frame-by-frame matching: SURE hits 97.0% average success on LIBERO with a fifth of the training steps
Synopsis
In robot manipulation demonstrations, one task often admits several feasible poses and execution speeds, yet pointwise supervision treats each demonstrated action as the single exact target. SURE adds a lightweight transposed-convolutional probabilistic action decoder to an OpenVLA-7B backbone to capture spatial uncertainty among neighboring action bins, and pairs it with Soft-DTW alignment supervision that tolerates local timing differences between demonstrations. On LIBERO it reaches 97.0% average success after 10k training steps, above the 94.3% of OpenVLA-OFT trained for 50k steps; it averages 72.2% on five RoboTwin 2.0 tasks and 57.5% on six real-world tasks.
Interpretation
Robot manipulation policies no longer have to match the demonstrator's exact pose and tempo frame by frame; actions are learned as a locally correlated probability distribution. Earlier VLA action heads mostly treated each demonstrated action as an isolated point target, and although diffusion- or flow-based heads can express multimodality, supervision was still pointwise, so the model needed more data to discover that neighboring actions deserve similar credit. On the four LIBERO task suites, SURE trained for 10k steps reaches 97.0% average success, against 87.1% for OpenVLA-OFT at 10k steps and 94.3% at 50k steps on the same backbone.
Spatial uncertainty is carried by a small probabilistic decoder and temporal uncertainty by a differentiable sequence-alignment loss, and together they reduce sensitivity to an individual demonstrator's habits. Transposed convolutions treat adjacent action bins as ordered neighbors rather than unrelated classes, while Soft-DTW turns the dynamic time warping used for speaking-rate differences in speech recognition into a differentiable soft minimum. On LIBERO-Spatial, removing both the probabilistic decoder and Soft-DTW drops success from 99.0% to 94.4%; removing only Soft-DTW gives 97.4%, and removing only the probabilistic decoder gives 96.8%.
The data-efficiency gain is measured separately under a fixed training budget across demonstration fractions, not only as a consequence of fewer training steps. With the budget fixed at 10k steps on LIBERO-Spatial and 10%, 25%, 50% and 100% of demonstrations, SURE improves success over OpenVLA-OFT by 21.4, 4.2, 4.2 and 4.6 percentage points respectively. With only 25% of demonstrations SURE reaches 95.0%, above the 94.4% OpenVLA-OFT obtains with the full dataset; with only 10% it reaches 93.8%, within 0.6 percentage points of full-data OpenVLA-OFT.
Perspective
The evaluation covers the four LIBERO task suites, five bimanual RoboTwin 2.0 tasks under the Easy setting with 100 rollouts per task, and six tabletop tasks on LeRobot SO-101 arms with 20 trials per task; the real-world demonstrations were collected by the authors acting as operators and involve objects rather than people.
On LIBERO, both OpenVLA-OFT and SURE are evaluated with a single seed, and the data-efficiency experiments are also all single-seed; the RoboTwin 2.0 baselines are taken from the RoboTwin 2.0 paper and from JEPA-VLA, while the real-world OpenVLA-OFT baseline was trained by the authors for 50k steps on the same mixed-task demonstrations. The authors state that the reported success rates do not measure operational safety and do not justify unsupervised use around people.
