Stories about VLA
3 related stories
SMILE: Smooth Motion for Improved Long-Horizon VLA Execution
AI InsightSMILE predicts B-spline coefficients and decodes them into smooth action sequences, changing only the action representation rather than the model architecture to extend fixed horizons for VLA models. Compared with prior approaches that relied on larger models or more inference calls to improve long-horizon accuracy, it improves accuracy and amortized inference efficiency across four baselines including SmolVLA and Evo1, with SMILE-Evo1 achieving 98.0% and 1.1x speedup on LIBERO. This suggests B-spline smoothed action representation could become a low-cost universal improvement for long-horizon VLA execution.Key TakeawayUnlike extending models or adding inference, SMILE replaces action representation with B-splines to boost long-horizon accuracy and speed.Why It MattersLong-horizon VLA degradation has often been blamed on model capacity; SMILE proves action representation is a key variable, yielding significant efficiency gains with near-zero architecture changes.Who's Affected- AI ResearchersGain a plug-in action representation method to improve long-horizon VLA without retraining backbone models.
- Robotics ResearchersReal-world experiments validate smooth action sequences for execution stability, transferable to robot control pipelines.
- Vla Model DevelopersCan directly replace the action decoding layer to improve benchmark scores and inference speed at low cost.
What's NextWatch whether SMILE becomes default in subsequent VLA work and whether B-spline representation generalizes to more complex action spaces and longer task chains.Importance 72/100DriftingVLA: Native One-Step Vision-Language-Action Generation via Per-Dimension Temporal Drifting
AI InsightDriftingVLA introduces a native one-step vision-language-action model, replacing iterative flow matching with a distribution-drifting objective to generate a complete action chunk in a single forward pass. Compared with conventional flow-based VLAs that require multi-step refinement, this method significantly reduces online control latency. Per-Dimension Temporal Drifting (PDTD) respects distinct control semantics across action dimensions, offering a new direction for real-time robot decision-making.Key TakeawayCompared with flow-based VLAs requiring multi-step refinement, this method generates a full action chunk in one forward pass.Why It MattersReducing latency in online robot control makes VLA more practical for real-time decision-making, potentially advancing embodied AI deployment.Who's Affected- AI ResearchersOffers a new one-step action generation paradigm; distribution-drifting objectives may extend to other generative tasks.
- Robotics DevelopersSingle-step inference reduces compute overhead, easing deployment on resource-constrained robots.
- Embodied AI IndustrySpeeds VLA transition from offline simulation to online real-world control with faster response.
What's NextWatch for latency benchmarks on real robots, quality of long action chunks, and PDTD's generalization across high-dimensional continuous actions.Importance 65/100Rethinking Language's Role in Efficient VLA for Autonomous Vehicles: Toward Smarter, Trustworthy Driving
AI InsightThis paper proposes the Language Residue taxonomy, categorizing language's role in VLA into train-time supervision, latent reasoning, conditional invocation, and full per-frame. The core shift is from 'whether to use language' to 'when and where to use language at inference,' as inference cost recurs per frame while training cost is paid once. This suggests onboard efficiency gains may come from inference strategy design rather than architecture.Key TakeawayShifts focus from language usage to when and where language is invoked at inference.Why It MattersGiven onboard compute limits, this taxonomy offers a systematic framework to cut VLA inference costs, potentially influencing deployment strategies.Who's Affected- AI ResearchersGain a new taxonomy for language's inference role in VLA, directly usable for further experiments.
- Autonomous Driving DevelopersCan use the taxonomy to evaluate latency-memory trade-offs of inference strategies in onboard systems.
- AutomakersSee a new direction for efficient VLA deployment, though real gains await engineering validation.
What's NextWatch for empirical models or AD deployments that adopt L3/L4 conditional invocation strategies.Importance 65/100