Robotics

LfD / PbD

1994ActivePublished: 31 July 2026Updated: 20 September 2026Published

Key innovation

Teaching a robot new tasks by having a human show them (demonstrations) rather than hand-coding - the robot acquires a skill/policy from examples of the task being performed.

How it works

1) A (human) teacher demonstrates the task: via kinesthetic guiding of the arm, teleoperation or an observed recording. 2) The system records the demonstrations - trajectories, forces, vision, robot state. 3) A policy/skill is extracted from the demonstrations (e.g. via imitation learning / behavior cloning). 4) The robot reproduces and generalizes the task; optionally it improves further (e.g. via additional corrective demonstrations or reinforcement learning).

Problem solved

Hand-coding a robot for each new task is slow, costly and needs experts. Teaching by demonstration lets a non-technical user 'show' the task, and the robot learns it from examples.

Components

Demonstration interfaceTeacher input

The way the task is shown: kinesthetic guiding, teleoperation or video observation.

Official

Demonstration recordingData collection

Capturing trajectories, forces, vision and robot state during the demo.

Official

Policy extractionSkill learning

Turning demonstrations into a policy/skill (e.g. imitation learning / BC).

Official

Reproduction and generalizationTask execution

The robot reproduces the learned task and generalizes it to new conditions.

Official

Implementation

Reference implementations

robomimic

Python · ARISE Initiative (Stanford)

LeRobot

Python · Hugging Face

Implementation pitfalls

Distribution shift (compounding errors)High

A policy learned from demonstrations can drift outside the demo distribution, compounding errors (the behavior-cloning problem).

Fix:Corrective demonstrations (DAgger), more/diverse data, combining with RL.

Human-robot correspondenceMedium

Different human and robot embodiment/kinematics makes transferring demonstrations hard (especially from observation).

Fix:Motion retargeting, teleoperation/kinesthetics instead of pure observation.

Cost and quality of demonstrationsMedium

Good demonstrations are labor-intensive; inconsistent or suboptimal demos degrade the learned policy.

Fix:Ergonomic teleoperation interfaces, quality control, data filtering.

Evolution

Original paper · 2009 · Robotics and Autonomous Systems (Elsevier), 2009 · Brenna D. Argall

A survey of robot learning from demonstration

Brenna D. Argall, Sonia Chernova, Manuela Veloso, Brett Browning

1994

Programming by Demonstration (early work)

Early work on programming robots by demonstration (teleoperation and guiding) in industrial and research robotics.

2004

Apprenticeship Learning via Inverse RL

Inflection point

Abbeel and Ng show learning a policy from demonstrations by recovering a reward function (inverse RL) — a key alternative to pure behavior cloning.

RL (concept)Apprenticeship Learning via Inverse Reinforcement Learning (Abbeel & Ng, 2004) (paper)

2009

LfD survey (Argall et al.)

Inflection point

Argall, Chernova, Veloso, Browning publish the canonical survey 'A survey of robot learning from demonstration', framing the field.

2011

DAgger — reducing distribution shift

Inflection point

Ross, Gordon and Bagnell introduce DAgger (Dataset Aggregation) — iterative aggregation of corrective data addressing compounding errors in behavior cloning.

Imitation Learning / Behavior Cloning (concept)A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning (Ross et al., 2011) (paper)

2023

Scaling demonstrations for foundation models

Teleoperation and demonstrations at scale feed robotic foundation models and skill libraries (embodied AI).

2023

Diffusion Policy — visuomotor learning from demonstrations

Chi et al. show diffusion policies learned from demonstrations that model multimodal action distributions — a new standard in manipulation learned from demos.

Diffusion Policy (concept)Diffusion Policy: Visuomotor Policy Learning via Action Diffusion (Chi et al., 2023) (paper)

Hyperparameters (configurable axes)

Demonstration modalityCritical

Kinesthetic guiding, teleoperation or observation (video).

Policy-learning methodHigh

E.g. behavior cloning, inverse RL, dynamic movement primitives.

Behavior Cloning (BC)Supervised learning of a state→action mapping from demonstrations.

Inverse RL (IRL)Recovering a reward function from demonstrations (apprenticeship learning).

DAggerIterative corrective data aggregation that mitigates distribution shift.

Dynamic Movement Primitives (DMP)Parametric movement primitives learned from single demonstrations.

Diffusion PolicyVisuomotor generative (diffusion) policy learned from demonstrations.

Number of demonstrationsHigh

How many demos are needed to learn and generalize the task.

Computational complexity

Time complexity: Nie dotyczy (paradygmat uczenia/HRI, nie algorytm). Space complexity: Nie dotyczy.

Compute bottleneck

Number and quality of demonstrations

Effectiveness depends primarily on demonstration quality and coverage (data), not raw compute; the bottleneck is costly collection of good demonstrations and out-of-distribution generalization.

Depends on

Jakosc/pokrycie demonstracjiKoszt pozyskania demonstracji

Execution paradigm

Primary mode

Conditional

This describes the phase of reproducing the learned policy; the teaching paradigm itself is not a compute paradigm.

Activation pattern

Input dependent

Routing mechanism

Parallelism

Parallelism level

Partially parallel

Demonstration collection is sequential (human), but policy training and multiple demo stations can run in parallel.

Scope

TrainingInference

Constraints

!Human demonstrations are collected sequentially; policy training can be parallelized.

Hardware requirements

Primary

A teaching/HRI paradigm - independent of accelerator type; kinesthetic demonstration benefits from back-drivable actuation, and policy training may use GPUs.