contextual and motion cues
Information derived from human actions that provides context and guidance for robot learning.
Category: Concepts
3D point-based policies
Policies that utilize 3D point data to inform the manipulation actions of a robot’s multi-fingered hands.
embodiment gap
The disparity between human and robot capabilities that complicates the transfer of learned behaviors.
AINA
AINA is a framework designed to learn robot manipulation policies from human demonstrations in natural environments.
Aria Gen 2 glasses
Lightweight smart glasses equipped with a high-resolution RGB camera and 3D pose estimation capabilities.
CUDAGraphs
A representation of a sequence of CUDA operations that can be executed on a GPU, allowing for efficient parallel processing.
Large Language Model
Large language models are advanced neural network architectures trained on vast amounts of text data to understand and generate human-like text.
Long-Tail Distribution
A statistical distribution where a small number of events or items are extremely common, while a large number are rare.
Adaptive Drafter
A lightweight draft model that is continuously trained on idle GPUs to align with the target model during long-tail response generation.
Adaptive Rollout Engine
A system that manages a memory-efficient pool of pre-captured CUDAGraphs and selects appropriate speculative decoding strategies for input batches.