Go2 Navigation & Inspection with RL, VLA & World Models
Master's Thesis | Universität Osnabrück | Ongoing | Deadline: 30 November 2026
What are we trying to do?
Train and evaluate three different AI approaches to control a Unitree Go2 for autonomous robot navigation and inspection inside a simulated industrial environment.
ISAAC SIM / ISAAC LAB
Camera Depth LiDAR IMU Robot State
\ | | | /
\ | | | /
└─────┴───────┴──────┴─────┘
│
▼
┌───────────────────┐
│ UNITREE GO2 │
└───────────────────┘
│
┌───────────────┼───────────────┐
▼ ▼ ▼
RL VLA WORLD MODEL
PPO OmniVLA V-JEPA / LeWM
│ │ │
└───────────────┼───────────────┘
▼
NAVIGATION + INSPECTION
│
▼
Compare Performance & Robustness
The robot should learn to move through an environment, avoid obstacles, locate a target, position itself correctly and support inspection tasks around industrial objects, machines or infrastructure.
The initial benchmark focuses on a controlled navigation task such as:
START
│
▼
Explore Environment
│
▼
Find Target
│
▼
Avoid Obstacles
│
▼
Navigate to Target
│
▼
Position Go2
│
▼
Inspect
│
▼
SUCCESS
Inputs
All approaches operate inside the same simulated environment.
| Input | Purpose |
|---|---|
| 📷 RGB Camera | Visual environment |
| 🌐 Depth Camera | Distance and geometry |
| 📡 LiDAR | Spatial surroundings |
| 🧭 IMU | Orientation and movement |
| 🦿 Joint States | Go2 body configuration |
| ⚙️ Actuator States | Robot movement state |
| 📍 Robot Pose | Position for training and evaluation |
| 💬 Language Goal | Semantic target for VLA |
| 🎯 Target State | Goal definition and evaluation |
The benchmark records these streams synchronously so that experiments can be reproduced and compared.
The Three Approaches
| Reinforcement Learning | VLA | World Model | |
|---|---|---|---|
| Approach | RL | Vision-Language-Action | World Model |
| Model direction | PPO | OmniVLA / VLA adaptation | V-JEPA 2 / LeWM |
| Main input | Sensors + robot state | Vision + instruction | Observation sequence |
| Output | Navigation action | Goal-conditioned action | Predicted representation / action support |
| Training | Environment interaction | Adaptation / fine-tuning | Recorded trajectories |
| Main question | Can it learn robust navigation? | Can it understand what to navigate to? | Can it predict what happens next? |
The exact VLA and World Model implementations remain subject to feasibility testing during development.
Shared Task
The important part of the thesis is that the approaches are not tested in completely different setups.
SAME ROBOT
Unitree Go2
│
▼
SAME ENVIRONMENT
Isaac Sim / Isaac Lab
│
▼
SAME TASK
Navigate + Inspect
│
▼
SAME SCENARIOS
│
▼
SAME EVALUATION PIPELINE
/ | \
/ | \
▼ ▼ ▼
RL VLA WM
This makes it possible to investigate where each approach performs well and where its limitations appear.
Simulation
┌──────────────────────────────────────────────┐
│ NVIDIA ISAAC SIM │
│ │
│ Environment Physics Sensors Go2 │
└──────────────────────┬───────────────────────┘
│
▼
┌──────────────────────────────────────────────┐
│ NVIDIA ISAAC LAB │
│ │
│ Training Reset Logic Rewards RL │
└──────────────────────┬───────────────────────┘
│
▼
TRAINING + BENCHMARK
The environment can be changed systematically to test whether a model really learned useful behavior.
Room Layout ░░░ → ▓▓▓
Lighting ☀ → ◐ → ◼
Obstacle Position A → B → C
Target Position 1 → 2 → 3
Sensor Noise Low → Medium → High
Target Appearance Known → Modified → Unseen
Hardware
The project uses two main GPU environments.
| Local Workstation | Remote Compute | |
|---|---|---|
| GPU | RTX 5090 | A100 |
| Isaac Sim | ✓ | |
| Isaac Lab | ✓ | |
| Rendering | ✓ | |
| Sensor Generation | ✓ | |
| RL Training | ✓ | ✓ |
| VLA Training | ✓ | |
| World Model Training | ✓ | |
| Large Batches | ✓ |
RTX 5090
│
├── Isaac Sim
├── Isaac Lab
├── Go2 Simulation
├── Sensors
└── Benchmark Execution
A100
│
├── VLA
├── World Models
├── Neural Training
└── Large Batch Experiments
Algorithms & Frameworks
| Area | Technology |
|---|---|
| Robot | Unitree Go2 |
| Simulator | NVIDIA Isaac Sim |
| Robot Learning | NVIDIA Isaac Lab |
| RL | PPO / RSL-RL |
| VLA | OmniVLA / VLA adaptation |
| World Models | V-JEPA 2 / LeWM direction |
| Deep Learning | PyTorch |
| Fast Simulation | MuJoCo / MJX |
| Local GPU | RTX 5090 |
| Training GPU | A100 |
| Project Management | Jira |
| Documentation | Confluence |
Benchmark Pipeline
SCENARIO
│
▼
ISAAC SIM / LAB
│
▼
GO2 + SENSORS
│
▼
┌─────────┬─────────┬─────────┐
│ RL │ VLA │ WM │
└────┬────┴────┬────┴────┬────┘
│ │ │
└─────────┼─────────┘
▼
ROBOT BEHAVIOR
│
▼
EPISODE LOG
│
▼
EVALUATION
│
▼
FINAL RESULTS
Each run stores the scenario, sensor observations, robot state, actions, target, model output, timing, collisions and final result.
What Do We Compare?
| Navigation | Robustness | Training | System |
|---|---|---|---|
| Success Rate | New Layouts | Training Data | Inference Time |
| Collisions | Sensor Noise | Training Time | GPU Usage |
| Time to Goal | New Targets | Sample Efficiency | Model Size |
| Path Efficiency | Obstacles | GPU Hours | Implementation Effort |
| Final Position | Lighting | Generalization | Reproducibility |
The result is a capability comparison, not one artificial overall score.
Project Management
The thesis is managed as a technical project using Jira + Confluence.
MASTER THESIS
│
┌────────────┴────────────┐
▼ ▼
JIRA CONFLUENCE
│ │
Work Packages Requirements
Tasks Architecture
Deadlines Research
Progress Decisions
Acceptance Criteria Experiments
Implementation Documentation
│ │
└────────────┬────────────┘
▼
TRACEABLE PROJECT
Jira
WP1 Setup
│
WP2 Research
│
WP3 Simulation
│
WP4 Benchmark + Data
│
WP5 Pilot Experiments
│
WP6 RL + VLA + WM
│
WP7 Final Benchmark
│
WP8 Thesis
│
▼
30 NOVEMBER 2026
The Jira workflow follows:
TO DO → IN PROGRESS → TESTING → DONE
Confluence
PROJECT
│
├── Charter & Scope
├── Requirements
├── Research
├── System Architecture
├── Experimental Design
├── Approach Specifications
├── Risk Register
├── Decision Log
├── Project Roadmap
└── Reproducibility
Jira manages what needs to be done.
Confluence documents what was designed, why decisions were made and how experiments are performed.
Current Status
JULY AUGUST SEPTEMBER
Setup ──────────► Simulation ─────────► Training
▲
│
WE ARE HERE
OCTOBER NOVEMBER
Training ─────► Benchmark ─────► Results ─────► Thesis
│
▼
30.11.2026
Current focus:
Go2 Environment
+
Sensors
+
RL / PPO
+
Reward Design
+
Episode Logging
↓
FIRST COMPLETE TRAINING PIPELINE
The next stage expands this pipeline toward the VLA and World Model implementations and then freezes the benchmark configuration for the final comparative experiments.
Final Goal
UNITREE GO2
│
┌────────────┼────────────┐
▼ ▼ ▼
PPO VLA WM
│ │ │
└────────────┼────────────┘
▼
SAME ROBOT TASK
▼
NAVIGATION + INSPECTION
▼
CONTROLLED BENCHMARK
▼
COMPARE RESULTS
▼
MASTER'S THESIS
▼
30.11.2026
The final deliverable is a working simulation demonstrator and reproducible benchmark showing how RL, VLA and World Model approaches can be trained and evaluated for autonomous robot navigation and inspection with a Unitree Go2.