Humanoid Robot Dancing and Running

1 September 2026

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan

Inclusive, Self-Reliant, Sustainable 

​

Universitas Gadjah Mada Team

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

Running

Result

The Methode

The Controller

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Methode: Deep Reinforcement Learning

• Classical control (DCM/DIP/CPG) handles nominal walking but can't adapt to disturbances or speed changes on its own
• We used policy-based RL: Proximal Policy Optimizatioin

WHAT is Deep Reinforcement Learning?

• A residual policy trained via reinforcement learning, layered on top of classical control (DCM/DIP/CPG)
• Learns to correct gait parameters rather than manual fine-tuned

​

WHY Deep Reinforcement Learning

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Architecture

Training hyperparameters

lr = 3×10⁻⁴, 4 epochs, 4 minibatches, gradient clip 0.5, horizon 64, 256 parallel environments.

Observation: 18 dimensions

Action: 6-dimensional residual

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

PPO Algorithm

Advantage is estimated with GAE(λ):

Clipped objective function:

Total Loss:

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Reward Design

Velocity tracking (exponential)

Flight-phase bonus

Energy penalty

Posture stability penalty

Total Reward per Steps:

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

Curriculum Learning & Training

Curriculum Learning

Promotion gate:

• fall rate ≤5%
• finish rate ≥95%
• tracking error <0.05 m/s

Training

PHASE 1: PRE-TRAINED WALK

PHASE 2: SPRINT CURICULLUM

• 100 updates
• completion 0.8% → context for Phase 2

• 4.003 m/s avg
• 256/256 finish

• 0% fall rate
• 5.16 s finish time

Finish Time: 120s

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Controller

Feedback-based PID Control

Central Pattern Generator - Gait Generator

DCM Capture Step

PPO Algorithm

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

Dancing

The Methode

The Controller

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Result: Tutting Dance

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Result: Tari Kecak Bali

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Result: Ateez San at City vs ATM

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Methode: PoseSeq

• Keyframe and Adjust ZMP
• Trial and Error with its controller

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

The Methode: PoseSeq

• Number of frames are exported into .csv
• Use it in controller options and use the controller module

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

DanceController: Motion Playback & Balance Control

Purpose

• Execute pre-generated humanoid dance motion in Choreonoid/AISTSimulator.
• Read joint trajectories from a CSV file.
• Generate joint position targets based on simulation time.
• Apply additional balance correction using IMU feedback

CSV Motion Trajectory ⟶ Time-Based Interpolation ⟶ Joint Position Reference ⟶ Balance Feedback ⟶ Corrected Joint Target ⟶ Humanoid Robot

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

CSV Loading & Joint Trajectory Interpolation

• Load Motion Data

​

​

​

​

​

​

​

• Linear Interpolation

The controller finds two neighboring motion frames and calculates the intermediate joint position

• Joint Command

The interpolated position is sent to each joint as its target position:

ioBody->joint(i)->q_target() = target;

​

​

​

Why Interpolation?

It prevents sudden jumps between motion frames and produces a smoother joint trajectory during simulation.

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

IMU-Based Balance Feedback

Sensors Used

• Accelerometer: estimates body pitch and roll tilt.
• Gyroscope: measures pitch and roll angular velocity.

​

The correction is limited to prevent excessive ankle movement and unstable feedback.

Tilit Estimation

pitchTilt = atan2(-a.x(), a.z());

rollTilt = atan2(a.y(), a.z());

​

PD Feedback

Correction=KP​×Tilt+KD​×AngularVelocity

Correction Limitation

maxCorr = 0.35 rad ≈ 20°

This is a simplified IMU-based ankle strategy, not a full ZMP preview controller. It is mainly intended to compensate for small balance disturbances

ugm.ac.id

Merakyat, Mandiri, Berkelanjutan | Inclusive, Self-Reliant, Sustainable 

​

“THANKYOU”

​

“TERIMAKASIH”