🤖 AI

VILAS: A VLA-Integrated Low-cost Architecture with Soft Grasping for Robotic Manipulation

The paper presents VILAS, a low-cost, modular robotic platform featuring a kirigami-based soft gripper and a unified communication framework, which successfully demonstrates the training and deployment of state-of-the-art vision-language-action models for safe, end-to-end manipulation of fragile objects like grapes.

Original authors: Zijian An (Luna), Hadi Khezam (Luna), Bill Cai (Luna), Ran Yang (Luna), Shijie Geng (Luna), Yiming Feng (Luna), Yue (Luna), Zheng, Lifeng Zhou

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zijian An (Luna), Hadi Khezam (Luna), Bill Cai (Luna), Ran Yang (Luna), Shijie Geng (Luna), Yiming Feng (Luna), Yue (Luna), Zheng, Lifeng Zhou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to teach a robot how to pick up delicate fruit, like grapes, without squishing them. Usually, building a robot smart enough to do this costs a fortune—think tens of thousands of dollars—and requires complex, expensive sensors to "feel" how hard it's squeezing.

The authors of this paper, VILAS, decided to build a cheaper, smarter, and more accessible version. They created a "low-cost" robot platform that can learn to do these tasks using a new type of AI called a VLA (Vision-Language-Action) model. Think of a VLA as a robot brain that can "see" the world, "read" a simple instruction (like "pick up the grape"), and immediately figure out how to move its arm to do it, all in one go.

Here is a breakdown of how they built it and what they found, using simple analogies:

1. The Robot Body: The "Budget-Friendly" Athlete

Instead of buying a $33,000 research robot (like the famous ALOHA system), they built their own using off-the-shelf parts that cost about $8,000 in total.

The Arm: They used a Fairino FR5, a collaborative robot arm. Think of this as a reliable, industrial-grade arm that is smooth and precise, but much cheaper than the custom ones usually used in labs.
The Hand: They used a standard electric gripper (the Jodell RG52-50), which is like a basic mechanical hand that can open and close.
The Eyes: They attached two cameras: one looking down from above (the "drone view") and one on the wrist (the "close-up view") to see exactly what the hand is touching.

2. The Secret Sauce: The "Origami" Soft Glove

The biggest challenge is picking up fragile things without crushing them. Usually, you need expensive sensors to measure force. VILAS didn't use those. Instead, they designed a soft, compliant extension made using a technique called Kirigami.

The Analogy: Imagine taking a flat sheet of paper and cutting specific patterns into it (like a stencil). When you squeeze this sheet, it doesn't just crumple; it naturally folds and curves into a 3D shape, like a flower blooming or a shell forming.
How it works: They 3D printed this "Kirigami shell" and attached it to the robot's hard gripper. When the robot closes its hand, this soft shell squishes and wraps around the grape gently. It acts like a passive safety net: the material itself absorbs the pressure, ensuring the grape isn't crushed, even if the robot doesn't have a "feeling" sensor. It's like wearing a thick, soft glove that protects your hand without you having to think about how hard you're squeezing.

3. The Brain: Teaching the Robot with "Teleoperation"

To teach the robot, they didn't write complex code. Instead, they used a teleoperation system.

The Analogy: Imagine a human wearing a "master" glove (a low-cost 3D-printed arm called GELLO). When the human moves their hand, the robot (the "follower") copies the movement instantly.
The human picks up grapes and puts them in a box 100 times. The robot records every move, every camera angle, and the voice command ("Pick up the grape"). This creates a "textbook" of 100 demonstrations.
They then took three different, cutting-edge AI brains ( $\pi_0$ , $\pi_0.5$ , and GR00T N1.6) and "fine-tuned" them using this textbook. They didn't teach the AI from scratch; they just showed it their specific examples so it could learn the task.

4. The Results: Who Won the Grape Game?

They tested the three AI models on a task: pick up three grapes in a row and put them in a box.

The "One-Shot" Champion ( $\pi_0.5$ ): This model was the best at picking up a single grape successfully (84% success rate). It was very good at the first attempt.
The "Marathon" Champion (GR00T N1.6): This model was the best at the whole sequence. While it was slightly less perfect on the very first pick (82%), it was much better at keeping going. It successfully picked up two grapes in a row 58% of the time, whereas the others dropped the ball after the first try.
The "Stutter" Problem: The other models tended to get confused after the first grape. They would go back to the same spot they just picked from, or drop the grape because their "grip" instructions got shaky. The GR00T model stayed focused and didn't repeat its mistakes as often.

5. The "Magic" Generalization

To see if the robots were truly smart or just memorized the grape, they swapped the grapes for cherries (which look different: smaller, redder, and smoother). They didn't retrain the robots; they just changed the voice command.

The Result: All three models could still pick up the cherries reasonably well. This proves the robots learned the concept of "grasping a small round object" rather than just memorizing the exact look of a grape.

Summary

The paper proves that you don't need a million-dollar lab to build a robot that can learn delicate tasks. By combining a cheap industrial arm, a 3D-printed "origami" soft glove, and modern AI models, they created a system that can learn to pick up fragile fruit. It shows that with the right "soft" mechanical design and smart AI, low-cost robots can do high-precision work.

Technical Summary: VILAS – A VLA-Integrated Low-cost Architecture with Soft Grasping for Robotic Manipulation

Problem Statement
Robotic manipulation of fragile objects, such as fresh agricultural produce, presents significant challenges due to the low tolerance of these items to excessive contact forces. Traditional approaches often rely on expensive multi-axis force-torque sensors, tactile arrays, or specialized impedance control systems, which increase hardware costs and system complexity. Furthermore, while recent advances in Vision-Language-Action (VLA) models offer a unified paradigm for end-to-end robotic control, training and deploying these models typically require high-cost, proprietary robotic ecosystems (e.g., ALOHA systems costing ~$33,000). This creates a substantial financial barrier for researchers aiming to develop and experiment with learning-based manipulation systems on accessible hardware.

Methodology
The authors present VILAS, a fully low-cost, modular robotic manipulation platform designed to support the complete VLA workflow—from teleoperation-based data collection to policy fine-tuning and real-time deployment.

Hardware Architecture: The system integrates three primary components:
- Manipulator: A Fairino FR5 collaborative arm (industrial-grade, brushless DC motors) chosen for its motion smoothness and precision at a cost of ~$4,000.
- End-Effector: A Jodell RG52-50 electric parallel gripper (~$1,500) augmented with a custom kirigami-based soft compliant extension. This extension, fabricated from PEBA using FDM 3D printing, induces predictable deformation under compressive loading. It provides passive mechanical compliance to absorb contact force variations, enabling safe grasping of fragile objects without explicit force sensing.
- Perception: A dual-camera setup comprising a base-mounted RealSense D455 (global workspace view) and a wrist-mounted RealSense D405 (close-range interaction view), totaling ~$1,000.
- Teleoperation: A GELLO leader arm (~$500) constructed from 3D-printed parts and Dynamixel servos, providing kinematically equivalent control for data collection.
- Total Cost: The complete platform costs approximately $8,000, significantly lower than comparable research systems.
Software and Control: The system utilizes a ZMQ-based communication architecture to coordinate the teleoperation client, data recorder, and robot actuators. During deployment, the control loop runs at 20 Hz. The system employs an action chunking strategy where a single model inference generates a sequence of actions (50 steps for $\pi_0$ / $\pi_{0.5}$ , 16 steps for GR00T N1.6), decoupling high-frequency control from the lower-frequency model inference.
VLA Evaluation: The authors fine-tuned and evaluated three state-of-the-art VLA models— $\pi_0$ , $\pi_{0.5}$ , and GR00T N1.6—on the VILAS platform. All models were initialized from publicly released pretrained checkpoints and fine-tuned on an identical dataset of 100 demonstration episodes collected via the VILAS teleoperation pipeline.

Key Contributions

System Design: The implementation of VILAS, a fully integrated, low-cost robotic platform that supports the end-to-end VLA workflow using modular, affordable components.
Soft Gripper Innovation: The design of a mechanically compliant, kirigami-based soft extension that facilitates safe manipulation of fragile objects without relying on expensive force sensors or complex force-control strategies.
Empirical Benchmarking: The deployment and comparative evaluation of three distinct VLA models ( $\pi_0$ , $\pi_{0.5}$ , GR00T N1.6) on identical low-cost hardware and demonstration data, providing practical insights into their deployment characteristics.

Results
Experiments were conducted on a grape grasping task (picking 3 grapes and placing them in a box) and a generalization test on cherries.

Single Grasp Success: $\pi_{0.5}$ achieved the highest single-grasp success rate (84%), followed by GR00T N1.6 (82%) and $\pi_0$ (70%).
Multi-Grasp Success: GR00T N1.6 demonstrated superior consistency in sequential tasks, achieving a 58% success rate for completing at least two consecutive grasps. This significantly outperformed $\pi_{0.5}$ (36%) and $\pi_0$ (22%).
Inference Latency: GR00T N1.6 exhibited the lowest mean inference latency (63.6 ms), though its per-step cost was higher due to a shorter action horizon (16 steps vs. 50 steps).
Generalization: All models demonstrated reasonable performance on cherries without additional fine-tuning, indicating that the learned policies capture manipulation primitives that generalize across visually distinct objects.
Failure Modes: Analysis revealed that $\pi_0$ and $\pi_{0.5}$ struggled with spatial attention updates (revisiting the same location) and gripper actuation consistency. GR00T N1.6 showed better consistency but exhibited a spatial bias toward searching near previous grasp locations.

Significance
The paper claims that VILAS validates the feasibility of training and deploying capable manipulation policies on low-cost, modular hardware. By demonstrating that state-of-the-art VLA models can be effectively fine-tuned and deployed on an ~$8,000 platform, the work lowers the financial barrier for robotic manipulation research. The results confirm that with appropriate mechanical design (soft compliant grippers) and system integration, high-performance VLA-based manipulation is accessible without reliance on expensive proprietary ecosystems or explicit force sensing. The authors position VILAS as a reproducible and accessible testbed for future VLA evaluation on real hardware.