{ "fetched_at": "2026-10-19T06:28:06Z", "source_url": "https://www.kaggle.com/competitions/bci-initiative-alvi-hci-challenge/data", "pages": [ { "name": "data-description", "content": "The dataset includes EMG forelimb muscle activity recordings and associated target hand pose (joint angles) from several healthy and amputee subjects. The data was acquired as subjects performed a cued hand movement task: a visual input showed the subject which movement to execute (e.g. flex index finger) and participants performed that movement with both hands. Movements varied in complexity, from single finger movement to more complicated gestures. Amputees attempted the imagined movement with the missing limb in parallel to the movement of the intact one. The pose of the hand opposite the arm wearing the EMG armband was reconstructed in 3D and is represented by 20 joint angle values. Your task is to train a model that can predict hand pose from the EMG data.\n\n### Dataset structure\nWhen you download the data (~5GB), you'll need to un-zip it. Within it, you'll find several folders for `health`y and `amputant` subjects and organized by the arm on which the EMG armband was worn (left/right). \nWithin each subject's folder, you'll find a folder called `preproc_angles` with the pre-processed data for the challenge. Each subject's data is further separated into `train` and `test` subsets (not all subjects have `test` data) for your convenience: you can use the `test` data to evaluate your model's performance on unseen data.\nEach `train/test` folder contains a series of `.npz` files, each storing data from a single movement (approximately 10 seconds). Each `.npz` file includes the EMG data and the joint angles data (see below).\n\nAlthough you can use any of the data provided to train your model, there's a specific subset of the data that is used to evaluate all models. This session belongs to one of the two amputant subjects, `fedya`, and is stored within a distinct subfolder called `submit`. \n\nRefer back to our [example code](https://github.com/BCI-I/BCI_ALVI_challenge/tree/main) to see how to load the data for your machine learning model and how to prepare your submissions!\n\n### Inputs\n\nThe inputs to your machine learning models are 8-dimensional time-series data, corresponding to the signals picked up by the 8 sensors in the Myo armband. The input data is already pre-processed to eliminate some noise and make the analysis simpler for you.\n\nNote that since subjects used the armband on either arm, the electrode position relative to the arm's muscles will be different. We provided some code for correcting for this difference while loading the data to simplify model training.\n\n\n### Targets\n\nThe output of your model should be a 20-dimensional time-series with the predicted angle for each hand joint at each time stamp. \n\n\n\n#### Columns\nEach `.npz` file in the dataset includes the following entries:\n\n- **myo_X**: with `X` a number in `[0,6]`. Time series data from each EMG sensor - the inputs to your model. \n- **ang_X**: with `X` a number in `[0, 19]`. Time series data with the ground truth values for each joint angle - what your model should predict. \n\n\n### The submission data\nAs described in the `Overview` tab, you'll use data from one of the two amputant subjects to prepare submissions to evaluate your model. This is the data in the `fedya_tropin_standart_elbow_left/submit` folder. The `npz` files in this folder **only include myo data** - the joint angles data is not provided such that models cannot be trained using this information. \n\nYou'll have to use these data and **only these data** to prepare your submission. We shared some code to show how to prepare your submissions. \n\n## A note on sampling\nThe data, both EMG and hand pose, is sampled at 200Hz. However, each model's predictions should be down sampled to 25Hz. When preparing your submissions, make sure that your model's predictions are down sampled accordingly. We share some code to illustrate this in detail. ", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "## Background\n\nThe brain controls movement by sending electrical signals through the nervous system to the muscles, causing them to contract. This contraction is the fundamental mechanism that generates movement in the body. When a muscle contracts, it pulls on the bone it is attached to, creating motion at the joint. This intricate process allows for a wide range of movements, from simple actions like lifting a cup to complex activities like playing a musical instrument. By measuring muscle activity, known as electromyography (EMG), we can gain insight into which muscles are contracting and thus infer the pose or position of a person's limbs.\n\n
\n \n
Courtesy of Scidraw.io, Smith Breault, Macauley.
\n
\n\n\nThis principle becomes particularly useful in the context of prosthetics for amputees. For an individual with an upper limb amputation, the remaining muscles in the residual limb still produce EMG signals when the person attempts to move their missing limb. By measuring these signals with dedicated sensors, we can predict what the pose of the hand \"should\" be, based on the muscle activity that is recorded. For instance, if the muscles that used to control the hand and wrist are still active, their EMG signals can indicate intended movements, such as grasping or pointing.\n\nBy applying advanced AI models to interpret these EMG signals, we can map the muscle activity onto specific hand poses. These AI models are crucial because they must accurately decode the complex patterns of muscle activity into precise movements. Once the intended pose is determined, this information can be used to send commands to a prosthetic hand, instructing it to take on the desired position. This process allows for natural and effective control of the prosthesis, enabling the user to perform a wide range of movements with their artificial limb.\n\nThe development of robust AI models is essential in this context. These models need to handle the variability in EMG signals and provide consistent, accurate mappings to ensure the prosthetic hand responds correctly to the user's intentions. With continued advancements in AI and sensor technology, we can enhance the functionality of prosthetic devices, offering amputees a more intuitive and seamless way to control their artificial limbs, thus improving their quality of life.\n\nOne of ALVI Labs' goals is to develop the next generation of AI models for this task. In this challenge, you get to have a go at it too!\n\n## This challenge\nWe've prepare some [example code](https://github.com/BCI-I/BCI_ALVI_challenge) to help illustrate how to load the data and train the baseline model. \n\n\n### The dataset\nALVI Labs generously shared some of the data they have acquired during their research. \nThis data was acquired using the [Myo Armbdand](https://www.researchgate.net/figure/The-Myo-Armband-Thalmic-Labs_fig2_346959518) from Thalmic labs to measure the activity of several forearm muscles while subjects performed several hand movements. The data includes muscle activity and reconstructed hand position in both healthy and amputee subjects.\n\nThe **goal** is to predict hand pose (expressed as the angle of multiple joints) from the measured muscle activation. Note that you will only be asked to make predictions for a single subject: data from the additional subjects is included in case it would help you engineer more effective models. Head over to the `Data` tab for more information about how the data is structured.\n\n### The Baseline Model\nALVI Labs prepare a model that can be used as baseline to assess the performance of your own!\nYou can see how exactly the model works [here](https://github.com/BCI-I/BCI_ALVI_challenge/blob/main/utils/hvatnet.py). Briefly, their model is based on the idea of a Convolutional Neural Network with a few additional modification. \n\nYou can use the baseline model as an inspiration for how to effectively tackle this challenge, but don't forget to explore other options too!\n\n### Prizes\nThe prizes offered as part of this competition are aligned to the BCI Initiative's goal to help promote research and education in this field. Winning teams will have a unique opportunity to have 1-on-1 meeting with some of the members of ALVI Labs' team to learn more about their experience launching this startup and what ALVI Labs' hopes to achieve in the future. In addition, outstanding teams will be offered the opportunity to work with the ALVI Labs' team to prepare an online blog post describing their approach to solving the challenge and, potentially, to work together on a conference paper submission. \n\n## Preparing for the competition\nAs the start date approaches, we will host additional material on this page, including example code and baseline models. We will also plan an online workshop, so stay tuned for more details!\n\n## Participating in person\nWhile this challenge is held asynchronously and online, we will endeavor to organize in-person periodic meetups prior to and while the challenge is running for those that want to meet with other participants. Stay tuned!\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Evaluation", "content": "Your goal is to predict joint angles as accurately as possible. For this reason, we will score prediction accuracy using MSE (Mean Squared Error) between the predicted angles (your model's output) and the ground truth data.\n\n## Submission File\nFor each sample in the test dataset, you will need to predict the angle of each 20 joints. Your submission should be a `.csv` that look like this:\n\n```\nsample_id,ang_0,ang_1,...,ang_19\n0,0.234.0.345, ..., 0.234\n1,0.652,-0.452, ..., 0.123\n2,-0.234,0.2342, ..., 0.465\n...\n```\n\nHead to the `Data` tab for more information about how the data is structured.", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "rules", "content": "* Anyone 16 years old or above is welcome to participate, regardless of their geographic location and educational or professional background. \n* Participants can compete individually or as part of a team (up to 6 people per team)\n* Participant must submit their solutions via the Kaggle platform, following Kaggle's guidelines. \n* Only the data provided can be used in training of ML models used for this challenge. \n* Model performance will be assessed using MSE on held-out data from a single object to determine winners. \n* The competition offers in-kind prizes, no cash prizes available. \n* Participants must adhere to Kaggle’s Code of Conduct.\n* Any form of cheating, including but not limited to, plagiarism, collusion, and use of unauthorized data, will result in disqualification.\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "abstract", "content": "Welcome to the BCI Initiative & ALVI Labs Challenge!\n\nThe BCI Initiative has partnered with ALVI Labs to bring you a unique opportunity to develop machine learning models that could one day power robotic prosthetic hands. Over a period of four weeks, participants from around the world will tackle this challenge, showcasing their skills and innovative solutions.\n\nIn this challenge, you will analyze electromyography (EMG) data from both healthy and amputee subjects, and your goal is to build a machine learning model that can predict the hand pose (the angle at which each joint is bent) from the muscle activation. This is the kind of model that can, one day, empower those living with upper limb amputations by allowing them to control robotic prosthetic hands!\n\nThe challenge is open to anyone, regardless of their geographic location or background. As the competition's start date approaches, we will be organizing online workshops to help you familiarize yourself with the data and we will share some code and baseline models to use as references.\n\nWinning teams will have the chance to collaborate with ALVI Labs in scientific research and publish the results of their work, contributing to advancements in rehabilitation and prosthesis control technologies.\nAbout the Organizers\n\n#### About the organizers\nThe [BCI Initiative](https://bci-i.github.io/) is dedicated to promoting education and research in Brain-Computer Interfaces (BCI) and Human-Computer Interaction (HCI) technologies. [ALVI Labs](https://alvilabs.com/) is a pioneering startup leveraging electromyography (EMG) and artificial intelligence (AI) to improve rehabilitation and prosthesis control. Join us in pushing the boundaries of neurotechnology and making a significant impact on the future of healthcare!", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }