Everyday Robot Dexterity
Robots that reliably grasp, fold and tidy in a kitchen, home or building site, where nothing is standardized.
Open in the interactive tree →In factories things sit in fixed places; in daily life they do not: laundry, dishes, cables and door handles are deformable, slippery and varied. Human hands with thousands of touch sensors are technically hard to replicate. This is known as Moravec’s paradox: what is easy for humans is hard for machines.
As of October 2026
According to the Stanford AI Index 2026, robots succeed in only about 12% of real household tasks, while AI agents reach 66.3% on the OSWorld computer-task benchmark. In 2026 humanoids mostly handle boxes and parts in fixed workflows (for example Figure’s BMW pilot, with above 99% placement accuracy according to the company), and home robots remain prototypes. Training data for movement is far scarcer than text data.
What is missing
- Large datasets for grasping and moving (there is no “internet of movements”)
- Robust, cheap hands with a sense of touch
- Reliability of 99.9% and more: one failure per thousand is too many at home
- Safe, lightweight actuators and batteries for close contact with people
- Transfer from simulation to the real world
Becomes possible once solved
- Robots in homes and care
- Automated construction sites, farming and skilled trades
- Flexible manufacturing of small batches
Open steps
- Movement data at scale Medium AI leverageCollect and generate the millions of robot-hours of grasping and manipulation data that language models got from the internet.
- Simulation to the real world High AI leverageTrain hands in simulation and transfer the policies to real robots without hand-tuning rewards and physics parameters.
- 99.9% reliability Medium AI leverageCut failure rates from percent to per-mille levels in unstructured homes, with recovery from rare mistakes.
- Touch-sensing hands Medium AI leverageMake cheap, durable robot hands with fingertip touch, and learn to use touch signals in contact-rich tasks.
- Safe, light actuators and batteries Low AI leverageBuild actuators and power packs that are light, back-drivable and safe near people, with hours of runtime.
Where AI could help
Medium AI leverage. AI is the main route to dexterity (learned policies, rewards, simulation), but missing movement data, touch-sensing hands and 99.9% reliability limit it.
- Train grasping and manipulation policies in large-scale simulation and transfer them to real robots
- Write and refine reward functions automatically instead of by hand
- Use vision-language-action models that generalize across tasks, scenes and robot bodies
- Generate synthetic training data and video world models for rare situations
Shown so far
- In October 2023 NVIDIA's Eureka used GPT-4 to write reward functions that beat human-engineered ones on 83% of 29 simulated robot tasks and taught a simulated hand to spin a pen. source
- In April 2025 Physical Intelligence reported (company paper) that its pi0.5 model cleaned kitchens and bedrooms in entirely new homes not seen in training. source
Prerequisites
- Deep Learning2012
- Humanoid Robots2026
Unlocks
- Household & Care Robots2030s?