You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$1,000 left to reach the funding goal.
I built ego2lerobot to turn first-person video into LeRobot datasets. Now I want to find out whether its hand-motion estimates help train a small forecasting model. I am asking for $1,000, mostly for GPU time. With $500, I can do the quality audit and a smaller pilot.
The code is Apache 2.0. The dataset card describes 648 episodes, 2.15 hours of video, factory-level splits and simple forecasting baselines. Those already exist. The full labels use CC BY-NC 4.0. This grant funds the next comparison.
Other researchers should be able to run these checks on their own footage and see where the data go wrong. I will share the evaluation code, settings and report, including failures. The funding supports this public research, not a private feature or product for Arete Robotics.
I will compare the data before and after filtering unreliable hand tracks. I will check visibility, missing detections and sudden motion changes, inspect a fixed sample, and keep the existing factory-level train/test splits.
I will compare simple baselines with one small model that also uses images. A GPU pilot will measure memory use and runtime. That will determine how many further runs fit the budget.
I plan six weeks after funding arrives and I have checked compatibility with existing support: two for the audit and pilot, three for comparisons, and one for releasing code and the report. This follows research I already plan to do; it adds no full-time appointment.
I will release the evaluation code, quality checks and error report so others can repeat the comparison. Results will include trajectory errors by data subset and examples of failures. Reference checks will stay separate from scores against automatic labels.
The filter might help or discard useful data. I will report either outcome. A forecasting gain concerns this task and dataset; it does not establish safe robot control or guarantee SOTA.
The code and scripts will be public. Data releases will follow source licences and access conditions, including restrictions on redistributing footage and labels.
My $1,000 budget is up to $850 for GPU rental, $100 for storage and transfer, and $50 for bank or currency-conversion costs. These are limits, not quotations or measured costs. I still need to choose a provider, check prices and run the pilot before larger experiments.
At $500, I would do the fixed audit, simple baselines and a smaller pilot: $400 for compute, $75 for storage and transfer, and $25 for payment costs. I will scale the work to the net funds received and handle unused money under the agreement.
I am requesting no salary, living costs or subscriptions. Paid experiments will wait until funding arrives.
I am Zhao Liu, a second-year PhD student at University College Dublin, working on this project independently. My experience includes stereo matching, multimodal 3D occupancy prediction and road-surface reconstruction. During internships at Horizon Robotics and PhiGent Robotics, I worked on autonomous-driving perception and failure cases.
I developed ego2lerobot through Arete Robotics, an unincorporated project. It tracks hands, estimates wrist motion and grip, can write descriptions using a local model, and exports LeRobot datasets. The human-motion estimates are approximate. Turning them into executable robot actions is a separate problem.
Project: https://arete-robotics.com/
Code: https://github.com/arete-robotics/ego2lerobot
Dataset card: https://huggingface.co/datasets/arete-robotics/egocentric-10k-actions
The labels may be too noisy, the filter may discard useful examples, or the image model may lose to a simple baseline. Compute could also limit the work. I will use the pilot to narrow the experiment before spending more and record failed runs.
I will publish the comparison even without a performance gain. If I cannot finish, I will notify the funder and handle unused money under the agreement. An already completed paper will not replace this work.
This project has received $0 in dedicated grants in the past 12 months. I have separate doctoral stipend and tuition support. This request covers additional experiment costs rather than another doctoral stipend.
Other applications for related research expenses and compute are pending, with no confirmed awards as of 25 October 2026. If one covers these costs, I will reduce or withdraw this request, or agree a separate budget with the funders. I will check compatibility before accepting money and never charge the same expense twice.