New Students on Sesame Street: What Order-Aware Matrix Embeddings Can Learn from BERT

Large-scale pretrained language models (PreLMs) are revolutionizing natural language processing across all benchmarks. However, their sheer size is prohibitive in low-resource or large-scale applications. While common approaches reduce the size of PreLMs via same-architecture distillation or pruning, we explore distilling PreLMs into more efficient order-aware embedding models. Our results on the GLUE benchmark show that embedding-centric students, which have learned from BERT, yield scores comparable to DistilBERT on QQP and RTE, often match or exceed the scores of ELMo, and only fall behind on detecting linguistic acceptability.

Traversing Steep and Granular Martian Analog Slopes With a Dynamic Quadrupedal Robot

Celestial bodies such as the Moon and Mars are mainly covered by loose, granular soil, a notoriously challenging terrain to traverse with (wheeled) robotic systems. Here, we present experimental work on traversing steep, granular slopes with the dynamically walking quadrupedal robot SpaceBok. To adapt to the challenging environment, we developed passive-adaptive planar feet and optimized grouser pads to reduce sinkage and increase traction on planar and inclined granular soil. Single-foot experiments revealed that a large surface area of 110cm2 per foot reduces sinkage to an acceptable level even on highly collapsible soil (ES-1). Implementing several 12mm grouser blades increases traction by 22% to 66% on granular media compared to grouser-less designs. Together with a terrain-adapting walking controller, we validate - for the first time - static and dynamic locomotion on Mars analog slopes of up to 25{\deg}(the maximum of the testbed). We evaluated the performance between point- and planar feet and static and dynamic gaits regarding stability (safety), velocity, and energy consumption. We show that dynamic gaits are energetically more efficient than static gaits but are riskier on steep slopes. Our tests also revealed that planar feet's energy consumption drastically increases when the slope inclination approaches the soil's angle of internal friction due to shearing. Point feet are less affected by slippage due to their excessive sinkage, but in turn, are prone to instabilities and tripping. We present and discuss safe and energy-efficient global path-planning strategies for accessing steep topography on Mars based on our findings.