QUADRUPED ROBOT

Legged Locomotion

Safe RL · Prediction · Sim-to-real

Quadruped Robot

We develop locomotion policies that recognize instability before failure and preserve agile, reliable motion when terrain and dynamics change.

Future-risk predictionDynamic locomotionSim-to-real

t+1…t+kpredictive safety horizon

2026current KSME study

RLstability-aware control

Activedeveloping program

Research question

Can the controller see a fall coming—and change course while recovery is still possible?

Quadruped robots can learn highly dynamic motion in simulation, yet a policy that only receives a penalty after falling has learned the consequence too late. Real hardware needs a controller that interprets body motion, contact, and command history as evidence about what is likely to happen next.

RV Lab is studying stability-aware reinforcement learning that predicts fall risk across a future horizon and feeds that estimate back into the present action. The program connects risk prediction, robust policy learning, domain randomization, and staged hardware validation.

Program pillars

Safety that participates in control

Stability is treated as a continuous, forward-looking signal—not a binary label observed after the episode ends.

01 / PREDICT

Preemptive fall awareness

Estimate how the current state and candidate action may evolve toward an unrecoverable posture before physical failure.

02 / LEARN

Safety-shaped policy learning

Use predicted risk during optimization so the controller learns recovery and task progress together.

03 / TRANSFER

Simulation-to-reality

Stress-test uncertainty, delay, friction, mass, and terrain variation before cautiously transferring policies to hardware.

Control loop

Predict, assess, and adapt

The future-risk model augments the locomotion loop without replacing the task policy.

01 · OBSERVE

Robot state

Joint motion, body velocity, contacts, commands, and terrain context.

02 · FORECAST

Future motion

Roll risk forward over a chosen prediction horizon from t+1 to t+k.

03 · EVALUATE

Stability risk

Distinguish recoverable motion from states that are converging toward failure.

04 · CONTROL

Safer action

Adapt foot placement and body motion while retaining commanded progress.

Current study

How far ahead should a robot predict?

A short horizon can miss slowly developing failure; a long horizon may become uncertain. The research examines that practical trade-off.

Quadruped robot stepping across uneven terrain

Stability-Aware RL · KSME 2026

Preemptive fall prediction for learned locomotion

The current paper introduces the predictive-risk concept and studies how the selected time horizon affects prediction and learning. It is an early program result, so this page deliberately separates proposed methodology from outcomes that still require full experimental publication.

Forecasts future risk rather than reacting only to terminal falls
Integrates safety information into policy optimization
Targets safer and more data-efficient hardware deployment

View official proceedings →

The available 2026 KSME item is a one-page conference abstract and does not provide a reusable full-scale figure or quantitative result table. The page therefore uses a program visual and labels the work as ongoing.

Evaluation roadmap

From a policy in simulation to a robot in the field

Evaluation is organized around the kinds of distribution shift that make learned locomotion brittle.

Terrain

Uneven ground

Slopes, steps, surface transitions, and incomplete foothold information.

Robotics operating in an outdoor agricultural environment

Field deployment

Outdoor mobility

Robust sensing and balance beyond controlled laboratory flooring.

Robot sensing in a safety monitoring application

Inspection

Persistent autonomy

Mobility, awareness, and safe recovery for long-duration inspection tasks.

Publication

Stability-aware reinforcement learning

The program currently has one verified proceedings paper, with expanded experimental work in progress.

KSME · 2026

Stability-Aware RL: Preemptive Fall Prediction for Quadruped Robots

A predictive-risk architecture for studying how forecast horizon influences fall prediction and policy learning.

View official proceedings →

Connected research

Mobility shares its foundations

Interested in safe learning for robots?

Join us in building predictive controllers, simulation environments, and careful sim-to-real validation workflows.

Meet the lab Browse publications