Explicit-Implicit Subgoal Planning for Long-Horizon Tasks With Sparse Rewards

The challenges inherent in long-horizon tasks in robotics persist due to the typical inefficient exploration and sparse rewards in traditional reinforcement learning approaches. To address these challenges, we have developed a novel algorithm, termed hlexplicit-implicit subgoal planning (EISP), designed to tackle long-horizon tasks through a divide-and-conquer approach. We utilize two primary criteria, feasibility and optimality, to ensure the quality of the generated subgoals. EISP consists of three components: a hybrid subgoal generator, a hindsight sampler, and a value selector. The hybrid subgoal generator uses an explicit model to infer subgoals and an implicit model to predict the final goal, inspired by way of human thinking that infers subgoals by using the current state and final goal as well as reason about the final goal conditioned on the current state and given subgoals. Additionally, the hindsight sampler selects valid subgoals from an offline dataset to enhance the feasibility of the generated subgoals. While the value selector utilizes the value function in reinforcement learning to filter the optimal subgoals from subgoal candidates. To validate our method, we conduct four long-horizon tasks in both simulation and the real world. The obtained quantitative and qualitative data indicate that our approach achieves promising performance compared to other baseline methods. These experimental results can be seen on the website https://sites.google.com/view/vaesi Note to Practitioners—This paper addresses the persistent challenges in executing long-horizon tasks in robotics, which are hindered by inefficient exploration and sparse rewards in traditional reinforcement learning methods. EISP overcomes these obstacles through the integration of a hybrid subgoal generator, a hindsight sampler, and a value selector, ensuring that the generated subgoals are both feasible and optimal, thereby facilitating the completion of the entire task by addressing each subgoal. Although preliminary results are promising, several challenges must be resolved in future research. Firstly, the automatic determination of the number of subgoals during training presents a challenge that may impact the low-level policy’s efficacy in achieving each subgoal. Secondly, the proposed method relies on an offline dataset that lacks successful trajectories from the initial state to the final state, complicating the exploration of subgoals towards the final goal. Finally, the scalability across multiple tasks is relatively limited due to the time-intensive nature of data collection for each task. Future research will focus on optimizing the allocation of timesteps to subgoals, enhancing exploration efficiency, and improving scalability across a broader range of tasks without extensive data collection.