Learning physical manipulation skills with simulators using realistic variations

About the project

The acquisition of manipulation skills in robotics involves the combination of object recognition, action-perception coupling and physical interaction with the environment. Several learning strategies have been proposed to acquire such skills. As for humans and other animals, the robot learner needs to be exposed to varied situations. It needs to try and refine the skill many times, and/or needs to observe several attempts of successful movements by others to adapt and generalize the learned skill to new situations. Such skill is not acquired in a single training cycle, motivating the need to compare, share and re-use the experiments.

In LEARN-REAL, we propose to learn manipulation skills through simulation for object, environment and robot, with an innovative toolset comprising: 1) a simulator with realistic rendering of variations allowing the creation of datasets and the evaluation of algorithms in new situations; 2) a virtual-reality interface to interact with the robots within their virtual environments, to teach robots object manipulation skills in multiple configurations of the environment; and 3) a web-based infrastructure for principled, reproducible and transparent benchmarking of learning algorithms for object recognition and manipulation by robots.

These features will extend existing softwares in several ways. 1) and 2) will capitalize on the widespread development of realistic simulators for the gaming industry and the associated low-cost virtual reality interfaces. 3) will harness the existing BEAT platform developed at Idiap, which will be extended to object recognition and manipulation by robots, including the handling of data, algorithms and benchmarking results. As use case, we will study the scenario of vegetable/fruit picking and sorting.

Work plan

3D graphics and physics engine
  • Selection of graphics and physics engine.
  • Rendering various forms of realistic variations for robot manipulation tasks.
  • Developing interfaces for researchers to link their robot models, algorithms with the simulator.
Reproducible research platform
  • Extending the existing BEAT platform to support workflows in robotics.
  • Definition of evaluation protocols, assessment metrics and data storage.
  • Developing evaluation benchmarks based on data produced in the context of LEARN-REAL.
Acquisition of manipulation data from demonstration
  • Learning from demonstration interface to help the robot acquire manipulation skills and create manipulation datasets.
  • Augmented reality (AR) or virtual reality (VR) headset to let users reconfigure objects and scenes in the virtual environment (scaffolding the environment), and learning scenario settings (e.g., modifying joints reconfiguration by robot guidance).
  • Evaluation baseline: kinesthetic teaching with Franka Emika's Panda robot equipped with a gripper and a RGB-D camera.
Generation of large-scale datasets
  • Toolset to generate synthetic RGB-D images of both textured and texture-less objects with a ground truth for scene understanding (i.e., objects, instances, and their 6D pose).
  • Generating grasp positions of synthetic objects, with objects isolated or superposed with a varying level of occlusion.
  • Use of multiple grippers, from simple parallel jaws to 3-finger hands.
Case study: recognition and manipulation of fruits/vegetables
  • Testing the developed toolchain within the case study of grasping fruits/vegetables.
  • Recognition of objects and segmentation of their instances, estimation of their 3D pose, and object grasping and manipulation by exploiting compact representations of object-adaptive movement primitives.
  • Domain adaptation to evaluate and correct the biases of synthetic data used for training w.r.t. real images observed by robots.