Operational definitions for AI data, alignment, agents, and physical AI.
17 resources
Agentic AI refers to AI systems that pursue goals through multi-step interaction with tools, software, environments, memory, or other agents, rather than producing only a single isolated response.
Data curation is the governed process of selecting, filtering, organizing, enriching, balancing, documenting, and versioning data so it is fit for a defined training, post-training, retrieval, or evaluation purpose.
Direct Preference Optimization (DPO) is a preference-based post-training method that optimizes a policy directly from preferred and dispreferred responses relative to a reference model, without first fitting a separate explicit reward model.
A golden trajectory is a reviewed reference path—or set of acceptable reference paths—showing how an agent can complete a defined task correctly under specified tools, permissions, policy, and environment state.
Inter-annotator agreement (IAA) measures how consistently two or more annotators assign labels, scores, spans, rankings, or other judgments to the same items under a defined annotation protocol.
LiDAR annotation is the creation or validation of labels on laser-scanned 3D point clouds or range data for detection, tracking, segmentation, mapping, localization, and scene understanding.
MCAP is an open, modular container file format for recording multiple channels of timestamped, pre-serialized data, commonly used for robotics and multimodal logs.
Model integrity is an operational umbrella term for evidence that an AI model or system behaves consistently with its specified purpose, constraints, provenance, security assumptions, and release requirements across its lifecycle.
Multimodal data combines two or more information modalities—such as text, image, video, audio, document layout, screen state, depth, point cloud, or sensor streams—in a record whose relationships matter to the target task.
AI red teaming is structured adversarial testing that probes a model or complete AI system for harmful, insecure, unreliable, policy-violating, or otherwise unacceptable behavior before and after deployment.
Reinforcement learning from human feedback (RLHF) is a family of post-training methods that uses human judgments to construct a learning signal for improving model behavior.
A ROS bag is a recorded collection of ROS messages and associated metadata used to capture, replay, inspect, and process communication from a Robot Operating System application.
Sensor fusion is the process of combining observations from multiple sensors or streams to estimate a state, detect an object, or make a decision more reliably than a single source alone.
Supervised fine-tuning (SFT) adapts a pretrained model by training it to reproduce curated target outputs for specified inputs or conversational contexts.
A tool-use trajectory is the ordered record of an agent’s interaction with tools and an environment from an initial task state to completion, failure, escalation, or termination.
A vision-language-action (VLA) model maps visual observations and language instructions to actions or action distributions for an embodied system.
A vision-language model (VLM) is a model that jointly processes visual inputs and language to represent, retrieve, generate, or reason about content across the two modalities.