In recent years, the surge in at-home fitness routines, especially during the global Covid-19 pandemic, has spotlighted a critical issue: improper exercise form leading to a significant rise in injuries. The U.S. Consumer Product Safety Commission reported a 48% increase in injuries related to at-home exercise during this period, underscoring the challenge many face without direct access to professional coaching. Addressing this gap, a pioneering team of researchers from Drexel University and Michigan State University has developed a cutting-edge prototype integrating artificial intelligence (AI), computer vision, and biomechanical modeling to offer real-time, precise exercise form coaching from streaming video footage.

This innovative program, dubbed BioCoach, marries advanced computer vision techniques with a vision-language model, allowing it not only to analyze human movement but also to generate live, anatomical feedback during various exercises. While numerous fitness coaching apps exist, few provide the specificity and immediacy of biomechanical correction delivered by a seasoned human trainer. BioCoach aims to bridge this divide by delivering targeted, timely cues rooted in the biomechanics of body motion, effectively emulating the nuanced guidance a knowledgeable coach would provide in person.

At the heart of BioCoach lies an intricate fusion of data processing algorithms. The system employs a dual-stream analysis approach: one stream utilizes a three-dimensional convolutional neural network (3D CNN) to capture visual appearance and motion dynamics, expertly recognizing distinct objects and movements within video sequences. Concurrently, a complementary stream estimates 3D skeletal posture and body morphology, extracting quantitative joint angles, ranges of motion, and exercise-phase data. This robust combination grants BioCoach an unprecedented depth of insight into the biomechanics underlying each repetition and posture captured on video.

The development team significantly enhanced the model’s training dataset by augmenting the Qualcomm Exercise Video Dataset (QEVD), a publicly available repository containing extensive exercise footage annotated with basic coaching feedback. Recognizing the sparse nature of original annotations, which often consisted of brief comments like “lower your body more,” the researchers re-annotated over 200 videos with detailed biomechanical targets and rationales. This enriched dataset included over 2,400 meticulously crafted notes specifying precise joint angles and motion thresholds, thus grounding the language model in authentic biomechanical context and timing.

This careful re-annotation process was integral not only in elevating the model’s linguistic precision but also in enabling rigorous evaluation of its feedback timing and relevance. By preserving the temporal alignment of coaching cues with specific exercise phases, the researchers ensured BioCoach’s ability to respond not just accurately but precisely when corrections are most beneficial—mirroring the instantaneous interventions of expert trainers.

BioCoach’s capacity to provide feedback is rooted in its ability to identify key joints relevant to individual exercises. For example, during squats, the system prioritizes analysis of the hips, knees, and ankles, while for push-ups, it focuses on the shoulders, elbows, and wrists. This targeted approach ensures that feedback remains specific and actionable, avoiding generic or irrelevant comments common in many current fitness apps. Additionally, by integrating detailed body shape and movement quality metrics, BioCoach can parse subtle deviations that might indicate compensatory patterns or strain risks.

The linguistic component of BioCoach translates intricate biomechanical data into natural language coaching cues with unparalleled clarity and relevance. Unlike more superficial feedback models, BioCoach articulates the significance behind each correction, explaining why a certain adjustment matters for distributing load or preventing injury. For instance, a suggestion might not only encourage “increasing elbow flexion to 90 degrees at the bottom of a push-up” but also clarify that “this adjustment helps distribute load evenly across joint structures,” thereby fostering user understanding and compliance.

In rigorous head-to-head testing, BioCoach was benchmarked against top-tier video-language AI models developed by prestigious institutions and corporations including MIT, NVIDIA, ByteDance, Alibaba, Salesforce, OpenAI, and leading Chinese universities. The evaluation involved feeding each program a combination of original QEVD videos and the newly annotated footage, assessing the response quality based on accuracy, anatomical correctness, detailed specificity, and timeliness.

The results were compelling. BioCoach outperformed its closest competitor, Stream-VLM (a collaboration between MIT and NVIDIA researchers) in text quality and relevance when evaluated on the original dataset. More strikingly, on the enriched dataset with biomechanics-based annotations, BioCoach demonstrated substantial gains across all metrics. Its feedback was notably more biomechanically accurate and rich with anatomy-specific details, establishing new standards for AI-driven exercise coaching.

The success of BioCoach highlights the profound benefit of integrating explicit 3D kinematic data and biomechanical constraints into AI coaching frameworks. By moving beyond mere pixel-level image analysis to structured, domain-specific knowledge, the system not only generates more accurate and insightful guidance but also becomes more interpretable and dependable, critical factors for user trust and safety in fitness applications.

Looking forward, the research team envisions expanding BioCoach’s capabilities to estimate joint reaction forces and muscle activation patterns from video input. Such enhancements would empower the system to detect even subtle compensatory movements or loading imbalances that can precipitate injury over time. These improvements could revolutionize both exercise and physical therapy by supporting users in receiving continuous, expert-level feedback remotely, effectively extending the reach of human trainers into digital spaces.

Dr. Feng Liu, assistant professor at Drexel’s College of Engineering and Computing and lead for the Visual Intelligence Lab, emphasized the transformative potential of BioCoach. “Our aspirations extend beyond simple encouragement,” he explained, “to actual biomechanically grounded coaching that helps individuals exercise safely and effectively. This integration of computer vision, 3D modeling, and language understanding is poised to redefine how AI supports human movement education.”

The development of BioCoach epitomizes a new wave of AI applications that intertwine deep learning and biomechanics, heralding an era where personalized, scientific exercise coaching is accessible anytime and anywhere. With ongoing refinement, such systems could democratize expert-level fitness guidance, mitigate injury risks, and ultimately promote healthier lifestyles across diverse populations worldwide.

Subject of Research: Not applicable
Article Title: From 3D Pose to Prose: Biomechanics-Grounded Vision–Language Coaching
News Publication Date: 27-Mar-2026
Web References: http://dx.doi.org/10.48550/arXiv.2603.26938
References: Feng Liu et al., arXiv preprint, 2026
Image Credits: Drexel University

Keywords: Artificial intelligence, Computer vision, Machine perception, Image processing, Natural language processing, Three dimensional modeling, Physical exercise

Would you take a payment to ramp down your electricity use? Would it change anything if you were doing so to help power a local data center?

Google just signed a new deal to help pay for a virtual power plant (VPP) in the largest power grid in the US. The agreement is with Voltus, a leading VPP and distributed energy resources platform.

Voltus will set up the virtual power plant, grouping together devices like electric vehicles and smart thermostats. It’ll pay customers to participate, and the company will dial back power or use the stored energy during times when the grid is stressed. Google will foot the bill for setting it up, and the extra capacity generated by the project will help run its data centers in the region.

This is one of the most concrete examples so far of a tech giant using a VPP to help meet energy demand for data centers. But there are still some lingering questions about just how far this sort of program can go, and what the limits are.

Last year, it felt as if everyone was talking about data center flexibility. A high-profile study from Duke University found that if data centers agreed to decrease their energy demand for roughly 40 hours per year, a whole bunch of them (about 100 gigawatts’ worth) could come online without making new power plants or transmission equipment necessary.

The underlying reason is that our power grid is designed not for our average energy use, but for the absolute maximum: the brutally hot July evening when everyone is blasting their air conditioners, watching Love Island, and microwaving popcorn. If a data center is willing to refrain from pulling so much power during those high-stress times, the grid can happily support it the rest of the year.

One lingering question here is about incentives: How would you get data centers to agree to this? After all, they might not have a very flexible load, especially now that AI use is more widespread—training a model can easily be delayed or shifted, but customer demand is more immediate. Giving up computing capacity could mean losing revenue.

Regulation is one approach that could work here. One proposal in the US would allow new data centers to come online years sooner if they agree to lower demand when the grid is nearing its max.  And a new Texas law requires large users to switch to backup power or curtail their demand in emergency situations.

Another approach is for data center operators to pay for other people to be flexible.

Voltus announced a new program in September that allows data centers to finance flexibility on their local grid. The company calls it “Bring your own capacity.” Google is now the first named customer taking advantage of this program.

In the new agreement, Voltus will pay people who agree to participate in the virtual power plant. The plant will be part of PJM, the grid that covers much of the US East Coast. The company says it will be able to aggregate up to 100 megawatts of distributed energy resources each year. The plant should be operational in 2027, according to Voltus.

This isn’t Google’s first foray into flexibility; the company has agreements with utilities across the US to limit or shift its own energy demand, which can help free up grid capacity. As the company pointed out in a blog post earlier this year, though, there are limits on how flexible a data center can be, and not every facility will be able to ramp down its power demand.

“There is no one solution for expanding grid capacity and we’re continuing to explore all options, including the many avenues for load flexibility,” said Michael Terrell, Google’s global head of advanced energy, in an emailed statement in response to written questions.

Once again, I’m wondering about incentives here. These companies are asking homes and businesses to be flexible. Will they agree?

A recent study in California looked at local people’s willingness to participate in managed electric-vehicle charging. Essentially, the program pays people to give up control of when they charge their EVs. This is another way to help smooth out electricity demand and ease the burden on the grid.

The problem? Not many people signed up. With no economic incentive, only 1% of EV owners enrolled in managed charging. At $40 per month (about 15% of their power bill), only 4.6% did.

This is a different situation and a different region from the one in which Google is working with Voltus. (It’s worth noting that the companies aren’t sharing how much they plan to pay the participants, which will obviously be a big determinant in participation for this kind of project.) 

But this study shows that even with money on the table, people may not always jump at the chance to cede control of their electricity demand. And it certainly feels relevant that about 70% of Americans oppose AI data centers in their area, according to recent Gallup polling. 

Being flexible sounds like a great idea in theory, and these financed VPPs could provide an immediate route to meeting energy demand. But as we move from idea to implementation, it’ll be interesting to see whether trial runs work as intended.  

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here. 

WhatsApp will charge businesses for using its AI agent based on token usage.