EveryDayVLA: A Vision-Language-Action Model for Affordable Robotic Manipulation (2511.05397v1)

Abstract: While Vision-Language-Action (VLA) models map visual inputs and language instructions directly to robot actions, they often rely on costly hardware and struggle in novel or cluttered scenes. We introduce EverydayVLA, a 6-DOF manipulator that can be assembled for under $300, capable of modest payloads and workspace. A single unified model jointly outputs discrete and continuous actions, and our adaptive-horizon ensemble monitors motion uncertainty to trigger on-the-fly re-planning for safe, reliable operation. On LIBERO, EverydayVLA matches state-of-the-art success rates, and in real-world tests it outperforms prior methods by 49% in-distribution and 34.9% out-of-distribution. By combining a state-of-the-art VLA with cost-effective hardware, EverydayVLA democratizes access to a robotic foundation model and paves the way for economical use in homes and research labs alike. Experiment videos and details: https://everydayvla.github.io/

• The paper presents a novel vision-language-action framework that jointly regresses discrete and continuous actions for affordable robotic manipulation.
• It introduces AdaHorizon, an adaptive ensembler that dynamically replans actions by measuring prediction disagreement, boosting success rates by up to 1.6%.
• Experimental results in simulation and real-world tasks demonstrate competitive performance, outpacing standard models while using a $300 hardware platform.

EveryDayVLA: Vision-Language-Action Models for Affordable Robotic Manipulation

Introduction and Motivation

The EveryDayVLA framework addresses persistent cost and accessibility barriers in robotic manipulation by integrating a state-of-the-art Vision-Language-Action (VLA) model with an open-source, low-cost ($\$300$) 6-DOF manipulator. Traditional robotic systems remain prohibitively expensive, and state-of-the-art VLA models exhibit brittleness in out-of-distribution scenarios, cluttered environments, or under severe distractors. EveryDayVLA combines collaborative training—jointly regressing discrete and continuous actions—and a novel adaptive horizon ensembler (AdaHorizon), enabling robust, efficient policy execution and dynamic uncertainty-driven replanning atop affordable hardware.

Figure 1: EveryDayVLA system overview: VLA model fine-tuned for a low-cost manipulator generates actions, which the AdaHorizon ensembler executes adaptively for reliable performance.

System Architecture

EveryDayVLA's pipeline ingests image observations together with natural language instructions, encodes them using pretrained visual (SigLIP, DinoV2) and language (Llama 2) backbones, and decodes outputs into both continuous and discrete action representations. Each representation predicts action "chunks" instead of single next actions, accelerating inference. The system architecture enables parallel decoding, leveraging the strengths of both modalities—precision from continuous outputs, semantic robustness from discrete tokens.

Figure 2: Architecture diagram: image and language instructions are encoded, sent to an LLM, and decoded into action chunks; AdaHorizon then determines execution length based on disagreement between the two.

Discrete actions are obtained by softmax over quantized bins; continuous actions via an MLP regression head. The loss weights (cross-entropy for discrete, L1 for continuous) are balanced for joint training, and action disagreement serves as a model uncertainty metric for adaptive replanning.

Adaptive Horizon Ensembler (AdaHorizon)

AdaHorizon dynamically determines the execution length for each action chunk by computing the mean absolute difference between the discrete and continuous action predictions. When disagreement exceeds a set threshold, replanning is triggered, ensuring high policy reliability and preventing command drift due to compounding autoregressive errors—a key problem in conventional VLAs. In practice, discrete actions are prioritized when their prediction aligns closely with the continuous output; this demonstrates an empirical advantage over fixed-threshold fusion schemes in simulation and real-world evaluation.

The ensembler maintains real-time constraints (minimum execution length, chunk-based replanning), and empirically improves success rates by up to 1.6% over the best competing ensembling strategies and pure modality baselines.

Hardware Design

The manipulator comprises off-the-shelf servo motors controlled by a PCA9685 PWM driver interfaced via I2C with an Arduino Uno, mounted on 3-D printed components and brackets. It provides a 0.2 kg payload, 382 mm reach, 0.7 m/s max speed, and $<$10 mm repeatability—all for $\sim$ \$300 per unit.

Training Paradigm

EveryDayVLA is fine-tuned on a custom dataset of 1,200 expert demonstrations, each pairing language instructions with RGB observations and end-effector poses, encompassing pick-and-place, manipulation, and stacking tasks with generic language templates. Training uses LoRA low-rank adaptation (rank=32) and mixed continuous/discrete objective, implemented atop the OpenVLA-OFT codebase. Both simulation (LIBERO benchmark, dual A100s) and real hardware (single A100) are supported with a streamlined teleoperation pipeline.

Experimental Results

Simulation (LIBERO Benchmark)

Across four LIBERO suites (Spatial, Object, Goal, Long), the AdaHorizon-equipped EveryDayVLA achieves average success rates trailing OpenVLA-OFT by only 3.9%, but matches or outperforms the best baseline on the Spatial suite (by +0.6%). AdaHorizon yields a consistent 0.8% performance boost over single-modality policies.

Inference rates reach up to 108.4 Hz (0.07 s latency), nearly matching the fastest baseline (OpenVLA-OFT), with minimal overhead from adaptive ensembling.

Real-World Evaluation

On in-distribution tabletop tasks (pick-and-place with blocks, balls, rocks under various instructions), EveryDayVLA outperforms prior models (OpenVLA, OpenVLA-OFT) by 49% on average. The hardware is robust across diverse objects, though delayed object release (timing misalignments) remains the primary failure mode.

Figure 3: Success rates for real-world pick-and-place tasks; EveryDayVLA substantial outperforms state-of-the-art VLAs for each object type and instruction.

Generalization and Robustness

On unseen tasks, environments, and with both static and dynamic distractors (clutter, moving humans), EveryDayVLA beats the next best method by 34.9%. In highly cluttered or dynamic scenes, average performance declines only $10$--$20\%$ relative to optimal, indicating strong robustness and superior generalization.

Figure 4: Performance under static and dynamic distractors: minimal drop-off observed compared to prior methods.

Action Ensembers Comparison

AdaHorizon surpasses temporal ensemblers (ACT), HybridVLA, and COGAct, with a 1.6% increase in success rate for fine-grained manipulation, demonstrating the value of dynamically tuned action chunk execution.

Practical Implications and Limitations

EveryDayVLA demonstrates that high success rates in both simulation and real-world environments can be achieved on drastically reduced hardware budgets. The combination of collaborative action prediction, uncertainty-driven replanning, and parallel action chunking significantly reduces inference latency and improves reliability. This enables affordable robotic foundation models and democratizes research and application for broader academic and hobbyist communities.

Limitations primarily stem from mechanical durability of the inexpensive servos, and sub-optimal performance in fine-dexterity tasks due to hardware precision constraints and limited number of expert trajectories for fine-tuning. Long-term hardware robustness and expanded expert data collection present clear directions for future work.

Theoretical and Future Outlook

The adaptive horizon framework points to broader applicability of dynamic action replanning and uncertainty quantification for sequential prediction in VLA and embodied AI domains. Hybrid decoders and collaborative losses can generalize to other multimodal robotic policy architectures. Improvements in tokenization, action chunking, and model-based planning may yield further gains in generalization and inference throughput, particularly under resource-constrained deployments.

A natural extension involves leveraging higher-fidelity actuators and larger, more diverse demonstration sets to further bridge the sim-to-real gap and address fine-grained manipulation. More sophisticated adaptive ensembling—potentially learning task- or context-specific replanning strategies—may also enhance generalization and robustness.

Conclusion

EveryDayVLA integrates a state-of-the-art vision-language-action policy with collaborative training and adaptive horizon execution atop highly accessible hardware, achieving competitive performance and robustness at a fraction of the prevailing cost. The adaptive horizon ensembler substantially enhances task success rates and model reliability under both standard and adversarial scenarios. While hardware durability and fine-dexterity precision remain areas for improvement, EveryDayVLA provides a tractable blueprint for accessible, scalable robotic foundation models and opens promising avenues for future research into uncertainty-aware, multimodal control.