BeyondContact: A Reinforcement Learning Framework for Airflow-Based Non-Contact Manipulation

Peiyu Yang, Mohamad S. Moudallal, Antonio Langella, Giulio Turrisi, Victor Barasuol, Claudio SeminiarXiv 2026 · arXiv

TL;DR

Researchers have developed a new reinforcement learning (RL) framework for airflow-based, non-contact robotic manipulation by creating a computationally efficient airflow extension for NVIDIA Isaac Lab. This framework enables robots to manipulate objects beyond their physical reach and precisely control their orientation using a custom-designed blower, achieving high success rates in zero-shot real-world deployments.

What Problem Does This Paper Address?

Traditional robotic manipulation relies almost entirely on physical contact—grasping, pushing, or pulling. While effective, this approach is strictly limited by the robot's kinematic workspace (how far it can reach) and the mechanical properties of the objects themselves (some things are too fragile or awkwardly shaped to touch).

Non-contact manipulation using airflow offers a compelling alternative: it can act on objects from a distance and even influence multiple objects at once. However, applying modern Reinforcement Learning to this domain has been difficult. High-fidelity fluid simulations (CFD) are far too slow for the millions of samples RL requires, while simpler "virtual particle" models often fail to capture the complex spatial decay and spreading of real airflow, leading to a "sim-to-real" gap where trained policies fail on actual hardware.

Methods

The authors bridge this gap with BeyondContact, a framework that combines efficient physics modeling with massively parallel simulation.

  1. Airflow Extension for Isaac Lab: They developed a robot-agnostic module that models airflow as a time-averaged, quasi-steady free turbulent round jet. Instead of simulating every air molecule, they use analytical models (like the "shifted inverse" model) to predict velocity fields.
  2. Surface-Integration Force Mapping: Unlike previous models that applied a single force vector to an object's center, this extension decomposes an object's surface into "primitives." It computes local forces on each primitive and aggregates them, allowing the simulation to capture complex eccentric torques and asymmetric effects.
  3. Mechatronic Design: To test the theory, they built a lightweight blower module using an electric ducted fan (EDF) and a 3D-printed enclosure, mounted alongside a standard gripper on an AgileX Piper-L robotic arm.
  4. RL Training: Using Proximal Policy Optimization (PPO) with an asymmetric actor-critic architecture, they trained two distinct policies:
  5. Hybrid Push–Blow (HPB): Coordinates physical pushing with airflow to move objects to targets far beyond the arm's reach.
  6. Rectangular Reorientation (RR): Uses airflow alone to rotate a cuboid to a specific target heading.
Fig. 1. Overview of the proposed reinforcement learning (RL) framework for airflow-based non-contact manipulation. A custom blower is integrated with an AgileX Piper-L robotic arm. The airflow simulation extension supports multiple object types and nozzle configurations, computing and applying airflow- induced forces and torques in simulation. The extension enables RL across multiple tasks, including hybrid push–blow (HPB) and rectangular reorientation (RR), with the trained policies subsequently deployed on the physical robot.

Datasets

The paper does not use a traditional pre-existing dataset. Instead, it relies on:

  • CFD Baselines: Data generated via OpenFOAM (a standard fluid dynamics tool) to validate their simplified analytical airflow models.
  • Real-world Measurements: Physical anemometer readings from their custom blower were used to identify model parameters (like decay coefficients and spreading rates) to ensure the simulation matched reality.

Key Insights and Ideas

  • Efficiency is King: By using analytical jet models instead of full fluid simulations, the authors achieved the high throughput necessary for RL (thousands of parallel environments) without sacrificing the physical accuracy needed for sim-to-real transfer.
  • Torque Matters: The "aha moment" for precision was moving from point-force models to surface integration. Capturing how airflow hits different parts of a rectangular block is what allowed the RR policy to achieve such fine orientation control.
  • Hybridization Extends Utility: Airflow isn't just a replacement for contact; it's an augment. The HPB policy learned to use contact to "set up" the object for a long-distance blow, effectively doubling the robot's functional workspace.
The force computation pipeline, showing how an object's surface is decomposed into primitives to calculate net force and torque.

Key Results

The framework demonstrated impressive "zero-shot" performance, meaning the policies worked on the real robot immediately after training in simulation:

  • Hybrid Push–Blow: Achieved an 80% success rate over 20 real-world trials, successfully moving objects to a target region 1.3 meters away—well beyond the arm's reach.
  • Rectangular Reorientation: Achieved a mean absolute heading error of just 2.6° across multiple target orientations using only airflow.
  • Simulation Performance: In the Isaac Lab environment, the policies reached success rates of 88% (HPB) and 92% (RR) after roughly 6 hours of training on a single modern GPU.

Limitations and Future Work

While successful, the authors noted that the robotic arm's own kinematic constraints sometimes limited the angles from which it could blow on an object, leading to failures in recovery.

Future work aims to:

  • Mount the system on a mobile base to further extend the operating range and provide more flexible blowing angles.
  • Test the framework on more complex objects with varied shapes and masses.
  • Explore multi-object manipulation scenarios where a single "gust" might be used to sort or aggregate items.

The project's code and materials are available at https://github.com/iit-DLSLab/BeyondContact.

Summary updated

Abstract

Contact-based robotic manipulation is constrained by end-effector reachability and the mechanical properties of objects. Airflow can exert forces without direct robot-object contact, enabling manipulation of objects beyond the reach or when direct contact is undesirable. However, the lack of suitable simulation has hindered the application of reinforcement learning to airflow-based manipulation, limiting the broader adoption of this mechanism compared with contact-based methods. To address this gap, we developed a robot- and task-agnostic airflow extension for Isaac Lab that supports massively parallel training. The extension uses experimentally identified airflow models to compute the forces and torques exerted on objects. Based on this extension, we train a hybrid push-blow (HPB) policy that combines contact-based pushing with airflow manipulation and a rectangular reorientation (RR) policy that controls a cuboid's heading using airflow alone. To enable physical execution, we design a lightweight, high-speed blower that coexists with a gripper. In zero-shot real-world deployment, HPB achieves an 80% success rate over 20 trials, while RR achieves a mean absolute heading error of 2.6 degrees across four target headings. The project materials are available at https://github.com/iit-DLSLab/BeyondContact.
BibTeX
@misc{yang2026beyondcontact,
  author = {Peiyu Yang and Mohamad S. Moudallal and Antonio Langella and Giulio Turrisi and Victor Barasuol and Claudio Semini},
  title = {{BeyondContact: A Reinforcement Learning Framework for Airflow-Based Non-Contact Manipulation}},
  year = {2026},
  eprint = {2611.03059},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.03059},
}
Figure