BoxDocker is an autonomous robotics spacecraft designed to combat the growing problem of space debris in Low Earth Orbit simulated in IsaacSim. The system is a versatile solution that uses a YOLOv11 deep learning model for real-time vision and a closed-loop control algorithm to detect and collect various space debris, as well as avoiding dangerous debris.
Space debris is one of the fastest-growing threats to space operations, with over 36,000 tracked objects larger than 10 cm and hundreds of millions of smaller fragments orbiting Earth at high velocities. Even a fleck of paint can cause severe damage to satellites or spacecraft due to the extreme kinetic energy involved. This accumulation of debris not only endangers current missions but also risks triggering the Kessler Syndrome — a cascade of collisions that could make Low Earth orbit increasingly hazardous. Addressing this problem requires innovative autonomous systems in space capable of detecting, avoiding, and removing debris of all sizes.
This project was run on Google Cloud Platform (GCP) and Azure.
-
Set up Cloud Storage
- Created a Google Cloud Storage bucket to hold all project files.
- Made three subfolders:
images→ pictures of debris from IsaacSimlogs→ records of spacecraft actionscheckpoints→ snapshots of the AI brain (YOLOv11) at different stages
-
Captured Simulation Images
- Ran IsaacSim and saved images showing:
- Small debris (1–10 cm)
- Medium debris (10–50 cm)
- Tiny flecks (dust or paint)
- Uploaded these images to Cloud Storage to train or test the AI.
- Ran IsaacSim and saved images showing:
-
Uploaded Files
- Uploaded all simulation images, logs, and AI checkpoints to their corresponding folders in Cloud Storage.
-
Launched Cloud Computer
- Started a Compute Engine VM with an NVIDIA GPU to run the AI and control program.
-
Installed Tools
- Installed Python and AI libraries on the VM to run YOLOv11 and the spacecraft control program.
-
Trained YOLOv11 (Optional / Verification)
- Used the simulation images stored in Cloud Storage to train YOLOv11.
- Saved progress as checkpoints.
- Tracked experiments and performance using Vertex AI.
-
Deployed the Control Program
- Loaded the spacecraft control program on the Compute Engine VM.
- The control program acted as BoxDocker’s “brain,” deciding:
- When to evade dangerous debris
- When to activate the correct capture mechanism
-
Connected AI to Control
- Used Pub/Sub to send real-time detection results from YOLOv11 to the control program on the VM.
-
Ran Simulation and Analysed Performance
- The control program moved the spacecraft in IsaacSim based on AI detections.
- Collected logs using Cloud Logging and analysed performance in BigQuery to check:
- Debris capture success
- Collision avoidance (>98% success)
- Time taken for each maneuver
Developed within the NVIDIA Isaac Sim simulation environment, the BoxDocker spacecraft is an example of such a system, specialized with multi-modal capture mechanisms. These include a prismatic docker for small debris (1-10cm), deployable doors for mid-sized objects (10cm+), and aerogel flaps for mcroscopic flecks (under 1cm). Currently there is no existing solution that tackles all these debris sizes making this system unique. This simulation-first approach allowed the rigorous testing of the system's ability to autonomously detect, evade, and capture a wide range of debris.
Our simulation results validate the technical feasibility of the BoxDocker. The YOLOv11 model achieved over 97% detection accuracy, while the control system maintained a debris evasion rate of over 98%. The capture mechanisms also demonstrated high success rates on their respective target types. These findings confirm that a single, scalable robotic solution for space debris removal is a viable concept.
Example Space debris being detected:
https://www.youtube.com/watch?v=KIuLJWdtXso
Due to working in partnership with Kings College I am not at liberty to share the source code.