Glossary

Large-data driven manipulation: Approach that uses large-scale human, robot, simulation, and operational data to train and continuously improve robot manipulation policies and control stacks.

Data flywheel: A loop where deployed systems collect demonstrations, failures, QA traces, operator overrides, and rework logs that improve the next model, controller, and process.

Embodiment gap: Mismatch between human hands, two-finger grippers, five-finger robot hands, and humanoids in degrees of freedom, force, sensing, viewpoint, and control interface.

Two-finger gripper: Parallel-jaw or pincer-like end-effector that is reliable and data-efficient for many industrial tasks but limited for in-hand manipulation.

Five-finger dexterous hand: High-DoF hand closer to human morphology; it aligns better with human hand data but increases sensing, control, calibration, and maintenance complexity.

VLA (Vision-Language-Action): Unified model that directly outputs robot actions from vision and language input.

ACT (Action Chunking with Transformers): Transformer-based action chunking — learns continuous action sequences from demonstrations to stabilize delayed-reward tasks.

Diffusion Policy: Policy learning via conditional denoising diffusion over action distributions.

UMI (Universal Manipulation Interface): Handheld gripper and camera demonstration interface for collecting in-the-wild manipulation data without operating a full robot.

EgoScale: NVIDIA GEAR-line research on scaling dexterous manipulation from large-scale egocentric human video pretraining.

Tactile/force-rich data: Manipulation data that includes contact locations, normal force, shear, force/torque, and slip events in addition to vision and proprioception.

Production readiness: Manufacturing adoption criteria beyond success rate, including cycle time, yield, scrap, MTBI, calibration drift, safety, and operator overrides.