RoboCurate: Harnessing Diversity with Action-Verified Neural Trajectory for Robot Learning
Abstract
RoboCurate enhances synthetic robot learning data by evaluating action quality through simulator replay consistency and augmenting observation diversity via image editing and video transfer techniques.
Synthetic data generated by video generative models has shown promise for robot learning as a scalable pipeline, but it often suffers from inconsistent action quality due to imperfectly generated videos. Recently, vision-language models (VLMs) have been leveraged to validate video quality, but they have limitations in distinguishing physically accurate videos and, even then, cannot directly evaluate the generated actions themselves. To tackle this issue, we introduce RoboCurate, a novel synthetic robot data generation framework that evaluates and filters the quality of annotated actions by comparing them with simulation replay. Specifically, RoboCurate replays the predicted actions in a simulator and assesses action quality by measuring the consistency of motion between the simulator rollout and the generated video. In addition, we unlock observation diversity beyond the available dataset via image-to-image editing and apply action-preserving video-to-video transfer to further augment appearance. We observe RoboCurate's generated data yield substantial relative improvements in success rates compared to using real data only, achieving +70.1% on GR-1 Tabletop (300 demos), +16.1% on DexMimicGen in the pre-training setup, and +179.9% in the challenging real-world ALLEX humanoid dexterous manipulation setting.
Community
๐ RoboCurate โ Harnessing Diversity with Action-Verified Neural Trajectory for Robot Learning
๐ Paper: https://arxiv.org/abs/2602.18742 ๏ฟผ
๐ What we do:
Propose RoboCurate, a synthetic robot data generation framework that improves neural trajectory quality through action-level verification and controlled diversity augmentation.
Introduce simulator-replay consistency filtering, which replays IDM-predicted actions in a simulator and evaluates motion alignment between generated videos and simulator rollouts using an attentive probe built on a frozen video encoder.
Expand observation diversity via a structured I2I (image-to-image) + V2V (video-to-video) pipeline, increasing scene and appearance variation while preserving action dynamics.
Apply a Best-of-N sampling strategy, where the learned action-consistency score serves as a critic to select the most reliable synthetic trajectory during generation.
Validate across large-scale simulation and real-world benchmarks (GR-1 Tabletop, DexMimicGen, and real ALLEX humanoid), achieving:
โข +70.1% relative improvement on GR-1 Tabletop (300 demos)
โข +16.1% on DexMimicGen (pre-training setup)
โข +179.9% on real-world ALLEX humanoid
โข Strong OOD gains (+162.3% on novel object tasks; emergent 0% โ 25% on novel behaviors)
๐ก Why it matters:
RoboCurate establishes an action-verified neural trajectory paradigm for synthetic robot data.
Rather than relying solely on video-level plausibility judgments from VLMs, RoboCurate directly evaluates whether predicted actions are physically consistent with observed motion through simulator grounding.
By jointly increasing visual diversity and enforcing action-level correctness, RoboCurate significantly enhances robustness, generalization, and real-world transfer of Vision-Language-Action models โ enabling more reliable policy learning from curated synthetic experience.
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper