Learning Pneumatic Non-Prehensile Manipulation With a Mobile Blower

We investigate pneumatic non-prehensile manipulation (i.e., blowing) as a means of efficiently moving scattered objects into target receptacle. Due to the chaotic nature aerodynamic forces, blowing controller must (i) continually adapt unexpected changes from its actions, (ii) maintain fine-grained control, since slightest misstep can result in large unintended consequences (e.g., scatter already pile), and (iii) infer long-range plans move robot strategic locations). tackle these challenges context deep reinforcement learning, introducing multi-frequency version spatial action maps framework. This allows for efficient learning vision-based policies that effectively combine high-level planning low-level closed-loop control dynamic mobile manipulation. Experiments show our system learns behaviors task, demonstrating particular achieves better downstream performance than pushing, improve over baselines. Moreover, we naturally encourages emergent specialization between different subpolicies spanning planning. On real equipped with miniature air blower, simulation-trained transfer well environment generalize novel objects.