Boston Dynamics' Robot Now Figures Out Vague Instructions on Its Own
Someday you'll just say 'tidy up' — and the robot will work out the rest.
Researchers built a system called CLUE — Closed-Loop contextual Uncertainty rEsolution — to help robots handle underspecified natural language tasks in unfamiliar environments. Instead of assuming goals are well-specified or that relevant information is provided upfront via a prior map, CLUE uses an LLM-derived policy to hypothesize task-relevant concepts and potential plans, then grounds those hypotheses into actions using a language-embedded map constructed online. It sequentially evaluates hypotheses through closed-loop environment interaction and refines its plans as it gathers new information. Deployed on a Boston Dynamics Spot across three real indoor and outdoor environments spanning 15 tasks requiring object disambiguation, functional inference, and occlusion reasoning, CLUE achieved a success rate within 7 percentage points of an oracle policy and outperformed an LLM-enabled planner without closed-loop feedback by a 4x margin. Supporting experiments showed that simply building and then querying a language-enriched map is insufficient to resolve complex contextual planning tasks — those approaches achieved roughly one third of CLUE's success rate while requiring over 10x more VLM tokens.
- A robot can now handle messy instructions like 'bring me the thing near the window' by testing guesses instead of failing outright.
- Tested on a Boston Dynamics Spot robot, it performed within 7 percentage points of a robot given perfect, cheat-sheet-level information.
- It used over 10 times less AI processing power than simpler approaches — meaning cheaper, faster, more practical robots.
Why It Matters
Vague human instructions won't stall robots much longer, which speeds up warehouse, hospital, and home helper robots.