FrameSkip: Learning from Fewer but More Informative Frames in VLA Training
Summary
FrameSkip is a data-layer frame selection method that improves Vision-Language-Action (VLA) policy training by prioritizing high-importance frames based on action variation and visual-coherence metrics, achieving a macro-average success rate of 76.15% across three benchmarks while using only 20% of unique frames.
View Cached Full Text
Cached at: 05/14/26, 08:17 AM
Paper page - FrameSkip: Learning from Fewer but More Informative Frames in VLA Training
Source: https://huggingface.co/papers/2605.13757 Published on May 13
·
Submitted byhttps://huggingface.co/VLyb
yubinon May 14
Authors:
,
,
,
,
,
,
,
,
,
Abstract
FrameSkip is a data-layer frame selection method that improves VLA policy training by prioritizing high-importance frames based on action variation and visual-coherence metrics.
Vision-Language-Action (VLA) policies are commonly trained from denserobot demonstration trajectories, often collected throughteleoperation, by sampling every recorded frame as if it provided equally useful supervision. We argue that this convention creates atemporal supervision imbalance: long low-change segments dominate the training stream, while manipulation-critical transitions such as alignment, contact, grasping, and release appear only sparsely. We introduceFrameSkip, a data-layer frame selection framework that scores trajectory frames usingaction variation,visual-action coherence,task-progress priors, andgripper-transition preservation, then remaps training samples toward high-importance frames under a target retention ratio. BecauseFrameSkipoperates only in thedataloader, it leaves the VLA architecture, action head, training objective, and inference procedure unchanged. Across RoboCasa-GR1, SimplerEnv, and LIBERO,FrameSkipimproves the success-retention trade-off over full-frame training and simpler frame selection variants, achieving amacro-average success rateof 76.15% across the three benchmarks compared with 66.50% for full-frame training while using a compressed trajectory view that retains 20% of unique frames in the main setting.
View arXiv pageView PDFProject pageGitHub0Add to collection
Get this paper in your agent:
hf papers read 2605\.13757
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.13757 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2605.13757 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.13757 in a Space README.md to link it from this page.
Collections including this paper1
Similar Articles
LiteFrame: Efficient Vision Encoders Unlock Frame Scaling in Video LLMs
LiteFrame proposes a lightweight video encoder with Compressed Token Distillation training that reduces latency and enables processing 8x more frames for long-form video understanding in Video LLMs, improving accuracy while reducing compute.
Frame selection is the whole game: notes on making LLMs watch video
Engineering notes on optimizing frame selection for feeding video to LLMs, covering scene detection, deduplication strategies, and token budget management.
PEEK: Picking Essential frames via Efficient Knowledge distillation
Introduces PEEK, an efficient dynamic frame sampling method that distills caption-conditioned frame relevance rankings from a teacher model into a lightweight temporal model, outperforming state-of-the-art methods in video captioning while maintaining computational efficiency.
SIEVE: Structure-Aware Data Selection for Imitation Learning with VLA Models
SIEVE is a structure-aware data selection method for vision-language-action imitation learning that identifies reusable visuo-motor primitives and transition interfaces, outperforming full-data training with only 50% of demonstrations and training steps.
PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models
PolicyTrim is a reinforcement learning-based post-training framework that improves action chunk utilization by 3× and reduces physical execution steps by 51.4% in Vision-Language-Action models, delivering up to 5.83× deployment speedup.