Learning from examples of success is an appealing approach to reinforcement learning that eliminates many of the disadvantages of using hand-crafted reward functions or full expert-demonstration trajectories, both of which can be difficult to acquire, biased, or suboptimal. However, learning from examples alone dramatically increases the exploration challenge, especially for complex tasks. This work introduces value-penalized auxiliary control from examples (VPACE); we significantly improve exploration in example-based control by adding scheduled auxiliary control and examples of auxiliary tasks. Furthermore, we identify a value-calibration problem, where policy value estimates can exceed their theoretical limits based on successful data. We resolve this problem, which is exacerbated by learning auxiliary tasks, through the addition of an above-success-level value penalty. Across three simulated and one real robotic manipulation environment, and 21 different main tasks, we show that our approach substantially improves learning efficiency. Videos, code, and datasets are available at https://papers.starslab.ca/vpace.

Abstract--Kinesthetic Teaching is a popular approach to collecting expert robotic demonstrations of contact-rich tasks for imitation learning (IL), but it typically only measures motion, ignoring the force placed on the environment by the robot. Furthermore, contact-rich tasks require accurate sensing of both reaching and touching, which can be difficult to provide with conventional sensing modalities. We address these challenges with a See-Through-your-Skin (STS) visuotactile sensor, using the sensor both (i) as a measurement tool to improve kinesthetic teaching, and (ii) as a policy input in contact-rich door manipulation tasks. An STS sensor can be switched between visual and tactile modes by leveraging a semi-transparent surface and controllable lighting, allowing for both pre-contact visual sensing and during-contact tactile sensing with a single sensor. First, we propose tactile force matching, a methodology that enables a robot to match forces read during kinesthetic teaching using tactile signals. Second, we develop a policy that controls STS mode switching, allowing a policy to learn the appropriate moment to switch an STS from its visual to its tactile mode. Finally, we study multiple observation configurations to compare and contrast the value of visual and tactile data from an STS with visual data Figure 1: Our STS sensor before and during contact with a cabinet knob from a wrist-mounted eye-in-hand camera. In visual mode, the camera sees through episodes from real-world manipulation experiments, we find that the gel and allows finding and reaching the knob, while tactile mode the inclusion of force matching raises average policy success rates provides contact-based feedback, via gel deformation and resultant by 62.5%, STS mode switching by 30.3%, and STS data as a dot displacement, upon initial contact and during opening. This dot policy input by 42.5%. Our results highlight the utility of seethrough displacement can also be used to measure a signal linearly related to tactile sensing for IL, both for data collection to allow force. Red circles highlight knob in sensor view.

Supervised imitation learning, also known as behavioral cloning, suffers from distribution drift leading to failures during policy execution. One approach to mitigate this issue is to allow an expert to correct the agent's actions during task execution, based on the expert's determination that the agent has reached a `point of no return.' The agent's policy is then retrained using this new corrective data. This approach alone can enable high-performance agents to be learned, but at a substantial cost: the expert must vigilantly observe execution until the policy reaches a specified level of success, and even at that point, there is no guarantee that the policy will always succeed. To address these limitations, we present FIRE (Failure Identification to Reduce Expert Burden in intervention-based learning), a system that can predict when a running policy will fail, halt its execution, and request a correction from the expert. Unlike existing approaches that learn only from expert data, our approach learns from both expert and non-expert data, akin to adversarial learning. We demonstrate experimentally for a series of challenging manipulation tasks that our method is able to recognize state-action pairs that lead to failures. This permits seamless integration into an intervention-based learning system, where we show an order-of-magnitude gain in sample efficiency compared with a state-of-the-art inverse reinforcement learning method and dramatically improved performance over an equivalent amount of data learned with behavioral cloning.

We present Learning to Place by Picking (LPP), a method capable of autonomously collecting demonstrations for a family of placing tasks in which objects must be manipulated to specific locations. With LPP, we approach the learning of robotic object placement policies by reversing the grasping process and exploiting the inherent symmetry of the pick and place problems. Specifically, we obtain placing demonstrations from a set of grasp sequences of objects that are initially located at their target placement locations. Our system is capable of collecting hundreds of demonstrations without human intervention by using a combination of tactile sensing and compliant control for grasps. We train a policy directly from visual observations through behaviour cloning, using the autonomously-collected demonstrations. By doing so, the policy can generalize to object placement scenarios outside of the training environment without privileged information (e.g., placing a plate picked up from a table and not at the original placement location). We validate our approach on home robotic scenarios that include dishwasher loading and table setting. Our approach yields robotic placing policies that outperform policies trained with kinesthetic teaching, both in terms of performance and data efficiency, while requiring no human supervision.

Adversarial imitation learning (AIL) has become a popular alternative to supervised imitation learning that reduces the distribution shift suffered by the latter. However, AIL requires effective exploration during an online reinforcement learning phase. In this work, we show that the standard, naive approach to exploration can manifest as a suboptimal local maximum if a policy learned with AIL sufficiently matches the expert distribution without fully learning the desired task. This can be particularly catastrophic for manipulation tasks, where the difference between an expert and a non-expert state-action pair is often subtle. We present Learning from Guided Play (LfGP), a framework in which we leverage expert demonstrations of multiple exploratory, auxiliary tasks in addition to a main task. The addition of these auxiliary tasks forces the agent to explore states and actions that standard AIL may learn to ignore. Additionally, this particular formulation allows for the reusability of expert data between main tasks. Our experimental results in a challenging multitask robotic manipulation domain indicate that LfGP significantly outperforms both AIL and behaviour cloning, while also being more expert sample efficient than these baselines. To explain this performance gap, we provide further analysis of a toy problem that highlights the coupling between a local maximum and poor exploration, and also visualize the differences between the learned models from AIL and LfGP.

Sequential modelling of high-dimensional data is an important problem that appears in many domains including model-based reinforcement learning and dynamics identification for control. Latent variable models applied to sequential data (i.e., latent dynamics models) have been shown to be a particularly effective probabilistic approach to solve this problem, especially when dealing with images. However, in many application areas (e.g., robotics), information from multiple sensing modalities is available -- existing latent dynamics methods have not yet been extended to effectively make use of such multimodal sequential data. Multimodal sensor streams can be correlated in a useful manner and often contain complementary information across modalities. In this work, we present a self-supervised generative modelling framework to jointly learn a probabilistic latent state representation of multimodal data and the respective dynamics. Using synthetic and real-world datasets from a multimodal robotic planar pushing task, we demonstrate that our approach leads to significant improvements in prediction and representation quality. Furthermore, we compare to the common learning baseline of concatenating each modality in the latent space and show that our principled probabilistic formulation performs better. Finally, despite being fully self-supervised, we demonstrate that our method is nearly as effective as an existing supervised approach that relies on ground truth labels.

Effective exploration continues to be a significant challenge that prevents the deployment of reinforcement learning for many physical systems. This is particularly true for systems with continuous and high-dimensional state and action spaces, such as robotic manipulators. The challenge is accentuated in the sparse rewards setting, where the low-level state information required for the design of dense rewards is unavailable. Adversarial imitation learning (AIL) can partially overcome this barrier by leveraging expert-generated demonstrations of optimal behaviour and providing, essentially, a replacement for dense reward information. Unfortunately, the availability of expert demonstrations does not necessarily improve an agent's capability to explore effectively and, as we empirically show, can lead to inefficient or stagnated learning. We present Learning from Guided Play (LfGP), a framework in which we leverage expert demonstrations of, in addition to a main task, multiple auxiliary tasks. Subsequently, a hierarchical model is used to learn each task reward and policy through a modified AIL procedure, in which exploration of all tasks is enforced via a scheduler composing different tasks together. This affords many benefits: learning efficiency is improved for main tasks with challenging bottleneck transitions, expert data becomes reusable between tasks, and transfer learning through the reuse of learned auxiliary task models becomes possible. Our experimental results in a challenging multitask robotic manipulation domain indicate that our method compares favourably to supervised imitation learning and to a state-of-the-art AIL method. Code is available at https://github.com/utiasSTARS/lfgp.