SplitNet: Sim2Sim and Task2Task Transfer for Embodied Visual Navigation

Abstract

We propose SplitNet, a method for decoupling visual perception and policy learning. By incorporating auxiliary tasks and selective learning of portions of the model, we explicitly decompose the learning objectives for visual navigation into perceiving the world and acting on that perception. We show dramatic improvements over baseline models on transferring between simulators, an encouraging step towards Sim2Real. Additionally, SplitNet generalizes better to unseen environments from the same simulator and transfers faster and more effectively to novel embodied navigation tasks. Further, given only a small sample from a target domain, SplitNet can match the performance of traditional end-to-end pipelines which receive the entire dataset 1 2 . * Work done during an internship at Facebook AI Research. 1 https://github.com/facebookresearch/splitnet 2 https://youtu.be/TJkZcsD2vrc

End-to-End Model-Free Reinforcement Learning for Urban Driving Using Implicit Affordances

“…Finally, there are two really recent articles closely related to ours. The first one by Gordon et al [10] introduced SplitNet on which they explicitly decompose the learning scheme in finding features from perception task and use these features as input to their model-free RL agent. But their scheme is applied to a completely different task, robot navigation and scene exploration.…”

Reduced Simulation: Real-to-Sim Approach toward Collision Detection in Narrowly Confined Environments

“…Thus, we should have pursued a more effective real-to-red.sim transfer. However, as indicated by [24], the separation of advanced visual perception and motion planning is advantageous for domain adaptation. For real-time MAV motion planning, where motion blurring is introduced, candidate applications are limited to non-agile and noncomplex trajectories.…”

“…Thus, it is not appropriate to use the existing sim-to-real type datasets and simulators or conventional methods directly for industrial applications in narrow or confined environments. Previous research [24] has handled this gap between real-worldbased simulators. This study has revealed the effectiveness of splitting visual perception and motion control.…”

“…Our proposed concept can also function as the methodology to make the learning of motion control independent from visual perception [24]. In other words, we required the real-to-red.sim transfer to be applicable with the trained deep model by reduced simulation into practical usage.…”

NeoNav: Improving the Generalization of Visual Navigation via Generating Next Expected Observations

“…TD-A3C, I2A and Gated-LSTM-A3C are all first trained via behavioral cloning using ground-truth paths. After pre-training, we update the three policy layers using a shaped reward based on the geodesic distance to the goal, geo(x, g), as described in (Gordon et al 2019): r t = geo(x t−1 , g)−geo(x t , g)+ζ,where ζ = −0.01 is a small constant time penalty. More implementation details are provided in the supplemental material.…”