A Solution to Adaptive Mobile Manipulator Throwing

Yang Liu, Aradhana Nayak, Aude Billard

TL;DR

This paper presents a novel framework for mobile manipulators to perform adaptive throwing tasks efficiently. By simplifying the problem to a planar one and combining machine learning for object dynamics with a robot kinematic feasibility model, they achieve throwing motion generation in under 1ms, enabling real-time adaptation to disturbances.

What Problem Does This Paper Address?

Automated throwing with mobile manipulators holds great promise for increasing flexibility and efficiency in tasks like logistics. However, a significant challenge lies in planning feasible and efficient throws that can adapt to a wide range of task specifications and unexpected disturbances. Existing methods often fall into two camps: model-based approaches that can be computationally expensive and difficult to generalize, or model-free methods that struggle with generalization and guaranteeing task success. This paper aims to bridge this gap by developing a method that is both computationally efficient and robust to dynamic changes.

Methods

The authors propose a framework that combines model-free learning with a model-based approach. Key to their method is:

  • Problem Simplification: The complex 3D throwing problem is reduced to a planar one, significantly cutting down computational costs.
  • Backward Reachable Tube (BRT): They model the set of valid throwing configurations by learning the object's inverted flying dynamics. This is achieved by sampling landing target sets and solving the flying dynamics backward in time, then representing the BRT using a neural network as a level-set function.
  • Velocity Hedgehog: To address robot kinematic feasibility, they develop a "velocity hedgehog" data structure. This discretizes the end-effector's height and throwing direction, storing the maximum feasible end-effector speed and the corresponding robot joint configuration for each cell. This allows for fast online queries of robot capabilities.
  • Initial Guess Generation: The BRT and velocity hedgehog are combined to generate initial guesses for the throwing configuration (joint angles, yaw, and pitch). These guesses are then refined using a reduced feasibility problem formulation.
  • Adaptive Throwing: The system's efficiency allows for replanning on the fly when disturbances occur. If the robot is perturbed during execution, it can quickly sample new throwing configurations and select an alternative, enabling adaptive behavior.

Datasets

The paper does not explicitly mention pre-existing datasets for training or evaluation. Instead, they generate their own data:

  • BRT Data: Approximately 75,000 throwing configurations are generated for the BRT by solving flying dynamics backward in time. An additional 80,000 configurations are generated as negative data (outside the BRT). This data is used to train a neural network for BRT representation.
  • Velocity Hedgehog Data: 1 million joint state samples are used to generate the velocity hedgehog.

Key Insights and Ideas

  • The "Tradeoff" between Yaw Angle and Throwing Velocity: The authors highlight a crucial insight: forcing the throwing yaw angle to zero (a common simplification) can filter out many optimal throwing configurations. They demonstrate that allowing the yaw angle to be a decision variable leads to more flexible and effective solutions.
  • BRT as a Feasibility Oracle: By learning the Backward Reachable Tube, the system gains an efficient way to query whether a given state is a valid throwing configuration, considering the object's dynamics.
  • Velocity Hedgehog for Efficient Kinematic Queries: This data structure provides a fast lookup for robot capabilities, enabling quick generation of initial guesses for joint configurations that can achieve desired velocities.
  • The Power of Offline Precomputation: The significant computational burden of generating the BRT and velocity hedgehog is shifted to an offline stage. This allows for extremely fast (sub-millisecond) online planning and adaptation.

Key Results

  • Computational Efficiency: The framework can generate throwing motions within 1 ms. The overall online computation time for a solution is approximately 500 µs.
  • High Success Rate: In simulation, the algorithm achieved a 99.4% success rate for throws, with failures primarily due to collisions (which are not explicitly handled by the core algorithm).
  • Adaptive Capability: The system demonstrated successful adaptation to disturbances, switching to a new valid throwing configuration within a reasonable time frame.
  • Generality: The method can handle arbitrary rigid objects (given their flying dynamics) and arbitrary target positions within the robot's kinematic limits.

Limitations and Future Work

  • Collision Avoidance: The current framework does not explicitly handle collision avoidance between the robot and the environment or the object and the environment during the throw. Failures in simulation were attributed to collisions.
  • Configuration Selection: The paper outputs batches of valid and feasible throwing configurations but does not specify a strategy for selecting the "best" one (e.g., time-optimal, risk-averse). This selection is considered out of scope for this work.
  • Modeling Error: The authors plan to investigate the influence of modeling errors in flying dynamics and trajectory tracking accuracy on real-world performance.
  • Real-World Implementation: The current validation is in simulation. The next step is to implement the algorithm on a real robot platform.

Summary updated

Abstract

Mobile manipulator throwing is a promising method to increase the flexibility and efficiency of dynamic manipulation in factories. Its major challenge is to efficiently plan a feasible throw under a wide set of task specifications. We show that the mobile manipulator throwing problem can be simplified to a planar problem, hence greatly reducing the computational costs. Using machine learning approaches, we build a model of the object's inverted flying dynamics and the robot's kinematic feasibility, which enables throwing motion generation within 1 ms for given query of target position. Thanks to the computational efficiency of our method, we show that the system is adaptive under disturbance, via replanning on the fly for alternative solutions, instead of sticking to the original throwing plan.
BibTeX
@inproceedings{liu2022solution,
  author = {Yang Liu and Aradhana Nayak and Aude Billard},
  title = {{A Solution to Adaptive Mobile Manipulator Throwing}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2022},
  doi = {10.1109/iros47612.2022.9981231},
}
Figure