Robotics

Robot learning

ActivePublished: 27 July 2026Updated: 27 July 2026Published

Key innovation

Replacing hand-programming of robots with machine learning: robots acquire skills (manipulation, locomotion) from data — demonstrations, experience or simulation — instead of hard-coded rules.

How it works

1) Data is collected: demonstrations (teleoperation, human video), interaction experience or simulation data. 2) A learning paradigm is chosen: imitation learning (behavior cloning), reinforcement learning (reward optimization) or self-supervised learning/pretraining. 3) A model/policy (e.g. VLA, VAM, diffusion policy) learns to map observations to robot actions. 4) The policy is evaluated and transferred to the robot (sim-to-real, fine-tuning on robot data), accounting for safety and generalization.

Problem solved

Hand-programming robots is labor-intensive, brittle and does not generalize to new tasks/environments. Robot learning lets robots acquire and generalize skills from data, coping with the complexity and variability of the real world.

Components

Policy / modelControl / decision

A model mapping observations (vision, sensors, state) to robot actions; e.g. a diffusion policy, a VLA or a VAM.

Official

Data sourceLearning signal

Demonstrations (teleoperation, human/egocentric video), interaction experience or simulation.

TeleoperationData collected on the robot by an operator.

SimulationSimulator data (cheap, scalable).

Official

Learning algorithmOptimization

The policy-optimization method: imitation learning, reinforcement learning or self-supervised learning.

Official

Sim-to-real transfer / deploymentDeployment

Transferring the learned policy to a real robot, bridging the sim-to-real gap and ensuring safety.

Official

Implementation

Reference implementations

LeRobot (Hugging Face)

Hugging Face

Implementation pitfalls

Sim-to-real gapHigh

Policies learned in simulation often fail on a real robot due to differences in dynamics and perception.

Fix:Domain randomization, fine-tuning on robot data, higher-fidelity simulation.

Low data efficiencyHigh

Collecting robot data is costly, and RL can be sample-inefficient.

Fix:Video pretraining, imitation learning, transfer and data sharing (multi-embodiment).

Safety and generalizationHigh

Learned policies may behave unpredictably out of the training distribution, which is risky for physical robots.

Fix:Safety constraints, human-in-the-loop, out-of-distribution testing.

Evolution

2016

End-to-end deep learning of visuomotor control

Inflection point

Deep learning begins mapping visual observations directly to robot actions, opening the modern robot-learning era.

2023

Vision-language-action (VLA) models

Robot foundation models (e.g. RT-2) combine web knowledge with robot action, improving generalization.

2024

Large-scale, multi-embodiment data and video-action models

Large-scale, multi-embodiment datasets and video-action (VAM) models accelerate efficient robot learning.

Hyperparameters (configurable axes)

Learning paradigmCritical

Imitation learning, reinforcement learning, self-supervised learning or hybrids.

imitation learningBehavior cloning from demonstrations.

reinforcement learningReward optimization.

Data sourceHigh

Teleoperation, human/egocentric video, simulation.

teleoperationOn-robot data.

simulationScalable, cheap.

EmbodimentMedium

Target robot platform (arm, humanoid, mobile).

Computational complexity

Time complexity: Zalezna od metody (IL/RL) i skali danych. Space complexity: O(model polityki + bufor danych/doswiadczenia).

Compute bottleneck

Data efficiency and the sim-to-real gap

Collecting robot data is costly, RL can be sample-inefficient, and policies learned in simulation do not always transfer to reality (sim-to-real gap).

Depends on

Ilosc/jakosc danychWiernosc symulacji