Research

Social robots or collaborative robots that have to interact with people in a reactive way are difficult to program. This difficulty stems from the different skills required by the programmer: to provide an engaging user experience the behavior must include a sense of aesthetics while robustly operating in a continuously changing environment. The Playful framework allows composing such dynamic behaviors using a basic set of action and perception primitives. Within this framework, a behavior is encoded as a list of declarative statements corresponding to high-level sensory-motor couplings. To facilitate non-expert users to program such behaviors, we propose a Learning from Demonstration (LfD) technique that maps motion capture of humans directly to a Playful script. The approach proceeds by identifying the sensory-motor couplings that are active at each step using the Viterbi path in a Hidden Markov Model (HMM). Given these activation patterns, binary classifiers called evaluations are trained to associate activations to sensory data. Modularity is increased by clustering the sensory-motor couplings, leading to a hierarchical tree structure. The novelty of the proposed approach is that the learned behavior is encoded not in terms of trajectories in a task space, but as couplings between sensory information and high-level motor actions. This provides advantages in terms of behavioral generalization and reactivity displayed by the robot. [Arxiv, Video]

Predicting human motion in unstructured and dynamic environments is challenging. Human behavior arises from complex sensory-motor couplings processes that can change drastically depending on environments or tasks. In order to alleviate this issue, we propose to encode the lower level aspects of human motion separately from the higher level geometrical aspects using data-driven dynamical models. In order to perform longer-term behavior predictions that account for variation in tasks and environments, we propose to make use of gradient-based constraint motion optimization. The present method is the first to our knowledge to combine motion optimization and data-driven dynamical models for human motion prediction. We present results on synthetic and motion capture data of upper body reaching movements (see Figure 1) that demonstrate the efficacy of the approach with respect to simple baselines often mentioned in prior work. [Humanoids-18]

We address the challenging problem of robotic grasping and manipulation in the presence of uncertainty. This uncertainty is due to noisy sensing, inaccurate models and hard-to-predict environment dynamics. We quantify the importance of continuous, real-time perception and its tight integration with reactive motion generation methods in dynamic manipulation scenarios. We compare three different systems that are instantiations of the most common architectures in the field: (i) a traditional sense-plan-act approach that is still widely used, (ii) a myopic controller that only reacts to local environment dynamics and (iii) a reactive planner that integrates feedback control and motion optimization. All architectures rely on the same components for real-time perception and reactive motion generation to allow a quantitative evaluation. We extensively evaluate the systems on a real robotic platform in four scenarios that exhibit either a challenging workspace geometry or a dynamic environment. We quantify the robustness and accuracy that is due to integrating real-time feedback at different time scales in a reactive motion generation system. We also report on the lessons learned for system building. [RAL-18, Video]

Motion optimization is a special brand of motion planning technique which generally consists of continuously bending the trajectory into a better solution. They are local methods as they seek to ameliorate the solution by descending the gradient of some objective function, and thus they typically suffer from getting trapped in local minima. However, by leveraging the optimization literature, and by pulling-back a model of the environment in the configuration space of the robot (in the Riemannian sense), it is now possible to solve motion planning problems with motion optimization for 7DoFs manipulators in cluttered workspaces within the order of a second and additionally perform continuous motion optimization with replanning steps under a second, allowing the behaviors demonstrated in the video. [IROS16, Presentation]

Inverse Reinforcement Learning (IRL) has been studied for more that 15 years and is of fundamental importance in robotics. It allows learning a utility function "explaining'' the behavior of an agent, and can thus be used for imitation or prediction of a given behavior by having solely access to demonstrated optimal or near optimal solutions. In this paper, we extend the IRL LEARning to SearCH (LEARCH) framework to train Convolutional Neural Networks (CNNs) using functional manifold projections, which we denote Deep-LEARCH. Earlier work on functional gradient approaches built large but flat additive models that continually grow in size. Our technique maintains the convergence advantage of functional gradient techniques (observed in linear spaces) while generalizing to fixed sized deep parametric models (CNN) by formally representing the function approximator as a non-linear sub-manifold of the space of all functions. We derive a simple step-project functional gradient descent method to walk across the manifold that is substantially more data efficient than traditional gradient steps consisting of a single back-propagation commonly used in Deep-IRL. We present preliminary experimental results showing higher-training rates on low-dimensional 2D synthetic data. We believe these ideas have broad implications for structured training beyond IRL as well as deep learning training in general. [WS-NIPS-16]

We have developed a framework that allows a human and a robot to perform simultaneous manipulation tasks safely in close proximity. The proposed framework is based on early prediction of the human's motion. The prediction system, which builds on previous work in the area of gesture recognition, generates a prediction of human workspace occupancy by computing the swept volume of learned human motion trajectories. The motion planner then plans robot trajectories that minimize a penetration cost in the human workspace occupancy while interleaving planning and execution. Multiple plans are computed in parallel, one for each robot task available at the current time, and the trajectory with the least cost is selected for execution. We test our framework in simulation using recorded human motions and a simulated PR2 robot. Our results show that our framework enables the robot to avoid the human while still accomplishing the robot's task, even in cases where the initial prediction of the human's motion is incorrect. We also show that taking into account the predicted human workspace occupancy in the robot's motion planner leads to safer and more efficient interactions between the user and the robot than only considering the human's current configuration. [IROS13, Video]

Pioneering work in motion planning with interaction constraints has been conducted at LAAS-CNRS, Toulouse, by Rachid Alami et al. The developed motion planning approach explicitly accounts for the human presence to synthesize friendly navigation and manipulation motions. This work, which was the first to the best of our knowledge to investigate a "planning'' approach to the problem of human-robot intelligent space sharing has been the basis of our doctoral work. We have extended the capabilities of this planner using sampling-based planning algorithms, which enable planning in the robot configuration space to find human-aware motions for high degree-of-freedom robots in cluttered environments. We have looked into combining global and local exploration of the high-dimensional cost landscape resulting from the mapping of elementary workspace cost functions in the configuration space of the robot. The framework makes use of the best of both approaches by finding an initial path using a global planner and then refining the solution locally with motion optimization. The final planner was applied to find legible motions for robot to human handovers. [Video, SORO14 , ICRA11, IARP10]