Intent Understanding

Figuring out what a person actually wants from a hint, without them naming the object or action directly.

Intent understanding is a robot's ability to infer what a person really wants when an instruction does not directly name the target object or action. An explicit instruction is something like “hand me the sponge”; an implicit one is more like “I spilled my drink, can you help?”, which requires common-sense reasoning: a spill needs wiping, and wiping needs a sponge. Google's 2022 SayCan used a large language model to break statements like this down into executable steps. CoRL 2024's Reasoning Grasping went further, mapping indirect instructions directly onto grasp poses — inferring which object to grasp and where to grip it. A 2025 evaluation, RoboBench, found that implicit instruction understanding remains a clear weak point for multimodal large models used as a robot's “brain.”

ExampleTold “I spilled my drink, can you help?”, SayCan plans: 1. locate the sponge, 2. pick up the sponge, 3. bring it to you, 4. done.

Also called

Implicit Instruction Following, Intention Reasoning, Implicit Instruction Understanding

Related

Instruction Following · Embodied Reasoning · SayCan · Vision-Language Model · LLM-based Task Planning · Human-Robot Interaction

Sources

SayCan: Do As I Can, Not As I Say (project page)
Reasoning Grasping via Multimodal Large Language Model
RoboBench: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models as Embodied Brain

As of

2025-10