AI Human Activity Recognition for On-board BAS Experiments
Indian Space Research Organisation(ISRO) · Miscellaneous · Software
The sequence-validation-with-voice-guidance framing is compelling and demos cleanly, but you build your own dataset so the model learns only your recorded procedure — nail the guidance-and-alerting loop, and be honest that the microgravity-orientation part, the space-relevant hard bit, is left optional for a reason.
Data: Requires synthetic/custom dataset generation by teams
What it actually is
On future space missions, communication delays make real-time ground support impossible, so an on-board AI could watch an astronaut perform a scientific experiment and check they are doing the steps right. The ask is a human-activity-recognition system that recognises and validates the sequence of a predefined experiment from fixed cameras, warns by voice if a step is skipped or out of order, and logs what was done.
What to build
An edge human-activity-recognition system that processes fixed-payload camera video locally to recognise the steps of a predefined experiment, using object detection, pose estimation and hand-object interaction; it tracks progress through the expected sequence, suggests the next step at the start of and after each step, raises a voice alert when a step is skipped or performed out of order, and generates a timestamped structured text log of what was conducted — optionally using orientation-agnostic 3D human mesh recovery so it works regardless of the astronaut's up-down orientation in microgravity, with teams generating their own custom training dataset.
Smallest thing that wins the room
Perform a short scripted multi-step procedure in front of a fixed camera and show the system tracking each step, suggesting the next one, raising a voice alert when you deliberately skip a step, and producing the timestamped log of what was done.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 34% of the 226 · #149 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: defence, intelligence and space bodies drew small fields.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
3/5The framing is compelling and the sequence-validation-with-guidance angle is genuinely useful, but the self-generated dataset means the model only learns the procedure you recorded, generalisation is weak from a small custom set, and the microgravity-orientation challenge that makes it space-relevant is exactly the hard part left optional.
Feasibility
2/5The component techniques exist, but the description requires teams to build their own dataset by recording a procedure, and training reliable step-recognition with pose and hand-object interaction from a small custom dataset is hard — the optional orientation-agnostic 3D human mesh recovery in microgravity is a research-grade addition on top.
Innovation scope
4/5Sequence validation of a procedure from activity recognition, with next-step guidance and out-of-order detection, is a genuinely interesting framing, and the orientation-agnostic tracking relative to the payload rack rather than the floor is a real open problem.
Clarity
4/5The description specifies the inputs, the required capabilities including next-step suggestion, skip and out-of-order voice alerts and the structured log, and marks the orientation-agnostic part optional, so the target is clear even though the dataset is self-generated.
Effort
MassiveGenerating a custom dataset, training object detection plus pose plus hand-object interaction, building the sequence-tracking and alerting logic, and running it at the edge is a large multi-part effort.
Demo-ability
EasyPerforming a procedure and watching the system guide, catch a skipped step by voice, and log the run is a concrete, self-explanatory demo you can script exactly.
In its favour
- Green flag: Scripting a procedure and catching a deliberately skipped step with a voice alert is a concrete, self-explanatory demo you fully control
- Green flag: Pose and hand-object interaction have mature open libraries like MediaPipe and MMPose to build on
- Green flag: The next-step-guidance and out-of-order-detection framing is genuinely useful beyond the space context
- Green flag: The structured timestamped log is a cheap, credible deliverable that a mission-support judge values
Against it
- Red flag: Teams generate their own dataset, so the model only learns the one procedure you recorded and generalisation from a small custom set is weak
- Red flag: The orientation-agnostic microgravity tracking that makes this genuinely space-relevant is research-grade and left optional, so most teams demo an Earth-oriented version
- Red flag: Reliable step recognition from pose and hand-object interaction is harder than it looks when steps are visually similar
- Red flag: An ISRO judge will ask how this transfers to actual microgravity conditions your Earth-recorded dataset cannot capture
What you will be writing
- Action / step recognition from video
- Pose estimation + hand-object interaction (MediaPipe / MMPose)
- Sequence tracking and out-of-order detection
- Custom dataset generation and labelling
- Orientation-agnostic 3D human mesh recovery (optional)
- Edge inference + voice alerting
- Human activity recognition
- Procedure validation
- Edge vision
Prior art to read before you start
experiment step sequence validation · procedure guidance from activity recognition · orientation-agnostic body tracking
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.