Toward Agile Maneuvers in Highly Constrained Spaces: Learning From Hallucination

Xiao, Xuesu; Liu, Bo; Warnell, Garrett; Stone, Peter

doi:10.1109/lra.2021.3058927

Cited by 34 publications

(26 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…LfH [9], [10] has been recently proposed to alleviate the difficulty of acquiring extensive or high-quality training data: from random exploration in a completely safe open space with complete safety, motion planners can be learned by synthetically projecting the most constrained [9] or augmented minimal [10] C-space onto the robot perception. Through carefully designed hallucination functions, these methods have shown fast and agile maneuvers on ground robots compared to classical motion planning and traditional learning approaches.…”

Section: B Machine Learning For Navigationmentioning

confidence: 99%

“…LfLH is tested on a ground and an aerial robot, both in simulated benchmark testbeds [13] and physical environments. Superior navigation performance is achieved compared to existing LfH approaches [9], [10] and classical sampling-based [14] and optimization-based [15] planners.…”

Section: Introductionmentioning

confidence: 96%

“…Learning from Hallucination (LfH) [9], [10] is a recently proposed paradigm to address the difficulty of 1) obtaining high-quality training data for traditional Imitation Learning (IL) from expert demonstrations [11] and 2) Reinforcement Learning (RL) from trial-and-error [12]. During LfH training, the robot executes a variety of random motion plans in a completely safe open space, imagines obstacle configurations for which the motion plans are optimal (called hallucination), and learns an end-to-end local planner as a mapping from the hallucinated obstacle configurations to the optimal motion plans in open space.…”

Section: Introductionmentioning

confidence: 99%

“…However, existing LfH methods require manually designed hallucination functions to generate the most constrained [9] or a minimal [10] obstacle set. While the former requires access to a fine-resolution global path and runtime hallucination during deployment, the latter assumes a short planning horizon to assure that one representative minimal unreachable set can approximately represent all possible minimal unreachable sets.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

From Agile Ground to Aerial Navigation: Learning from Learned Hallucination

Wang¹,

Xiao²,

Nettekoven³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

This paper presents a self-supervised Learning from Learned Hallucination (LfLH) method to learn fast and reactive motion planners for ground and aerial robots to navigate through highly constrained environments. The recent Learning from Hallucination (LfH) paradigm for autonomous navigation executes motion plans by random exploration in completely safe obstacle-free spaces, uses hand-crafted hallucination techniques to add imaginary obstacles to the robot's perception, and then learns motion planners to navigate in realistic, highly-constrained, dangerous spaces. However, current handcrafted hallucination techniques need to be tailored for specific robot types (e.g., a differential drive ground vehicle), and use approximations heavily dependent on certain assumptions (e.g., a short planning horizon). In this work, instead of manually designing hallucination functions, LfLH learns to hallucinate obstacle configurations, where the motion plans from random exploration in open space are optimal, in a self-supervised manner. LfLH is robust to different robot types and does not make assumptions about the planning horizon. Evaluated in both simulated and physical environments with a ground and an aerial robot, LfLH outperforms or performs comparably to previous hallucination approaches, along with sampling-and optimization-based classical methods.

show abstract

Section: B Machine Learning For Navigationmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 96%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

From Agile Ground to Aerial Navigation: Learning from Learned Hallucination

Wang¹,

Xiao²,

Nettekoven³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

show abstract

“…More importantly, end-to-end learning is extremely data-hungry, usually requiring millions of training data or training steps, and forgoes verifiable guarantees such as safety and explainability. On the other hand, other work which targeted a specific navigational component [11], [12], [13] has achieved superior performance when being compared to their classical counterparts. Therefore, it is promising to use machine learning at the subsystem or component level and to combine it with the structure of classical approaches [6].…”

Section: Arxiv:210507620v1 [Csro] 17 May 2021mentioning

confidence: 99%

APPLD: Adaptive Planner Parameter Learning From Demonstration

Xiao

Liu

Warnell

et al. 2020

IEEE Robot. Autom. Lett.

Self Cite

View full text Add to dashboard Cite

While current autonomous navigation systems allow robots to successfully drive themselves from one point to another in specific environments, they typically require extensive manual parameter re-tuning by human robotics experts in order to function in new environments. Furthermore, even for just one complex environment, a single set of fine-tuned parameters may not work well in different regions of that environment. These problems prohibit reliable mobile robot deployment by nonexpert users. As a remedy, we propose Adaptive Planner Parameter Learning (APPL), a machine learning framework that can leverage non-expert human interaction via several modalitiesincluding teleoperated demonstrations, corrective interventions, and evaluative feedback-and also unsupervised reinforcement learning to learn a parameter policy that can dynamically adjust the parameters of classical navigation systems in response to changes in the environment. APPL inherits safety and explainability from classical navigation systems while also enjoying the benefits of machine learning, i.e., the ability to adapt and improve from experience. We present a suite of individual APPL methods and also a unifying cycle-of-learning scheme that combines all the proposed methods in a framework that can improve navigation performance through continual, iterative human interaction and simulation training.

show abstract