BeyondContact: A Reinforcement Learning Framework for Airflow-Based Non-Contact Manipulation
TL;DR An Isaac Lab extension that applies forces from a measured blower airflow model, used to train RL policies for an arm with a fan mounted under its gripper. Zero-shot on the real arm, a push-then-blow policy moves a cube to a target beyond reach in 16 of 20 trials, and an airflow-only policy turns a cuboid to target headings with 2.6° mean error.
What they do
Airflow can push objects that an arm cannot reach, or should not touch. RL for this has been held back by the lack of a fast airflow simulator. BeyondContact adds airflow forces to Isaac Lab using a jet model fitted to measurements. It uses this to train two policies for an AgileX Piper-L arm carrying a small fan below its gripper, then deploys both zero-shot.
How it works
- Airflow model. The blower is modeled as a turbulent round jet: centerline velocity decays with distance (exponential or shifted-inverse fit) and spreads radially as a Gaussian whose width grows linearly. All parameters are identified from anemometer measurements and compared with an OpenFOAM CFD simulation.
- Isaac Lab extension. Object surfaces are split into primitives. Each primitive facing the jet gets a drag force from the local airflow velocity, and these are summed into a force and a torque about the center of mass. A reaction force is also applied to the link carrying the blower.
- Hardware. A 40 mm ducted fan in a 3D-printed housing with a magnetic, swappable nozzle, driven over PWM.
- Policies. PPO with an asymmetric actor-critic, 4096 parallel environments, curricula and randomized object, joint and airflow parameters. Hybrid push-blow (HPB) moves a cube to a target region; rectangular reorientation (RR) turns a cuboid with airflow only.
Results
- With the nozzle, the shifted-inverse fit has 0.90 m/s RMSE on centerline velocity; radial Gaussian fits reach R² ≥ 0.960.
- Training takes 5.7 h (HPB) and 6.6 h (RR) on one RTX 5090. Success in simulation is 88% and 92%.
- Real HPB with a 20 cm target 1.3 m from the base: 16/20 successes. The policy pushes the cube first, then blows it in.
- Real RR reaches 61.6°, 84.5°, 123.2° and 134.9° for targets of 60°, 90°, 120° and 135° (2.6° mean absolute error).
Limitations
HPB failures came from overly aggressive pushes and arm oscillation while blowing. The arm’s kinematics limit where air can be aimed from. Tests use a cube and a cuboid tracked by AprilTags.
Figures
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.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.