GSoC 2026: Applying for Project 20 - AVX2-Accelerated Collision Detection for Real-Time MPC in Humanoid Robotics - Prerequisite Task Request

[MDerazNasr]

Hi @RoboticsYY and @xwu2intel,

My name is Mohamed, and I am a first-year MS Computer Science student (Machine Learning Track) at Georgia Tech. My work focuses on ML systems and inference optimization. I have built a C++/CUDA inference engine for diffusion models from scratch, implemented INT8 quantization pipelines, and written GPU-parallel systems for my research under Dr. Adibi at GT.

I am applying for Project 20: AVX2-Accelerated Collision Detection for Real-Time MPC in Humanoid Robotics. The problem is clear to me: scalar collision detection is the bottleneck in the MPC loop and AVX2 batched distance queries are the fix. My CUDA kernel work gives me a solid base for thinking about instruction-level parallelism and I want to apply that to x86 SIMD in a real robotics control system.

I checked the Good First Issues board but all tickets in the Contributors Needed column are currently assigned. Could you assign me a prerequisite task directly? I am ready to start immediately.

Thank you for your time.

Mohamed Deraz Nasr
https://github.com/MDerazNasr/

[RoboticsYY]

Hi @MDerazNasr,

Thank you for your interest in our project. It is great to hear you have experience of creating diffusion transformers from scratch and making low level optimization. This project may not directly involve the AI model, but for a module post-processing trajectories output from humanoid VLA model, which requiring AVX SIMD acceleration to optimize model-predictive control (MPC) collision detection. There is already some discussions here with another applicant. You can find some links and guidance to ramp up quickly about the idea.

Please also note the upcoming deadline and ensure you submit an official application proposal through the program channel, as contributor ranking and final decisions will be based on that.

March 16 - 31 18:00 UTC: Potential contributors submit proposals through the program website for you to give feedback on.

April 1-20: Orgs review GSoC contributor proposals, determining how many to accept based on mentor availability, and rank them accordingly.

April 21 18:00 UTC: Deadline to submit ranked contributor slot requests by Org Admins

April 21 - 28: Google Program Admins review and assign org slots

April 29: Organizations receive notification of their accepted GSoC 2026 Contributors

April 30 18:00 UTC: Accepted GSoC 2026 GSoC Contributor projects are announced

In addition, another mentor of this project is @AnnikaWU. @AnnikaWU Please no hesitate to jump in and comment.

[MDerazNasr]

Hi @RoboticsYY, @xwu2intel, and @AnnikaWU,

Thank you for your response and for pointing me to the other discussion thread. I've gone through it and it gave me a lot of useful context.

I understand that the project scope is to accelerate OCS2's SelfCollisionConstraint using AVX2 SIMD, targeting the check-pair matrix that causes latency spikes in the MPC compute cycle. I've cloned both the OCS2 and VAMP repositories and am in the process of getting the environment set up on Ubuntu 22.04 with ROS2. My plan is to run the ocs2_mobile_manipulator example and establish a baseline profiling benchmark to measure the average and max MPC/MRT latency with collision detection on vs off.

I also took a look at VAMP's AVX2 implementation to understand how it batches sphere-pair checks using vector.hh and avx.hh, which is a good reference for what we'd want to replicate in OCS2's self-collision pipeline.

I am currently working on my proposal submission, but I was also wondering if I am still required to submit a PR to have my proposal considered. If so, would you be able to point me to a small task or area in OCS2 where I could submit a PR? If not I can continue working on my proposal as discussed.

Thank you again for your time!
Mohamed

[AnnikaWU]

Hi @MDerazNasr,

Thanks for your interest in our project, this is a good starting point.

Regarding your request: you don’t need to complete a prerequisite task or submit a PR in advance to apply for this project. The proposal itself is sufficient at this stage.

You can refer to these papers for a deeper understanding of VAMP and its collision algorithm CAPT:

• Motions in Microseconds via Vectorized Sampling-Based Planning
• Collision-Affording Point Trees: SIMD-Amenable Nearest Neighbors for Fast Collision Checking

Looking forward to your proposal.

Best regards,
Annika