River Islands High School — Class of 2029
Vyom Aggarwal
Student researcher at MIT and UCSC — building machines that reason, and doing the math underneath.
01 · Profile
I work where ML meets hardware.
I'm a sophomore at River Islands High School, dual-enrolled at San Joaquin Delta College. I specialize in the integration of AI and machine learning into engineering disciplines — mechatronics and aerospace in particular — along with the mathematics, simulation, and control that has to hold underneath for that integration to mean anything.
I like problems that stay interesting after the first correct answer: Is this number real, or is my instrument lying to me? Which assumption breaks first? What does the system do the moment it does?
That's why I work where ML meets hardware. A model that scores well on a benchmark and a model you would trust to run a physical system are not the same object, and the distance between them is where the engineering actually lives.
Quick spec
- Age
- 15
- Year
- Sophomore
- School
- River Islands HS
- Fields
- Mechatronics/ML
- Location
- Lathrop, CA
02 · Current research · UC Santa Cruz AIEA Lab
Teaching a broken robot to keep walking.
Fault Recovery in RL Locomotion Through Online Residual Adaptation — can a small correction module, updated during execution, restore a quadruped’s gait after a hardware fault without retraining the policy underneath?
01 — Premise
Policies are trained on robots that never break.
In the real world motors weaken, joints seize, and sensors start reporting values that aren’t true. A locomotion policy trained only under ideal conditions has never once been shown what failure feels like — so it has no response to it.
02 — Apparatus
A quadruped you can break on purpose.
Twelve revolute joints in PyBullet, physics at 240 Hz stepped down to 60 Hz control. PPO trained across ten seeds at 1.5M timesteps — six converged to a usable gait, and only those six are allowed into the experiment.
03 — Injection
At step 200, something fails.
Five fault families on a matched seed set, so every fault type is compared against identical initial conditions. Three corrupt the physics — what the robot can do. Two corrupt the observations — what the robot knows.
04 — Method
Correct around the fault, don’t retrain through it.
A small residual module sits between the frozen base policy and the actuators, updating online. The seized joint stays seized — the correction redistributes the gait around it. Measured against two bounds: no adaptation, and full retraining.
Not built yet — this is the hypothesis the apparatus exists to test.
05 — Measurement
First, build an instrument you can trust.
The original recovery criterion latched onto the first instantaneous velocity crossing and scored ~100% recovery on every fault — including 100% recovery alongside a 36.5% fall rate, which cannot both be true. It was rebuilt: velocity smoothed over a 30-step causal window, degradation has to be observed before recovery can be, the speed must hold for 30 consecutive steps, and a trial that falls is never scored as recovered.
Every recovery number predating that change was discarded.
06 — Finding
The faults separate cleanly.
Unadapted recovery runs from 0% on actuation_delay to 86.8% on sensor_noise. Every actuation-side fault is less recoverable than both sensor-side faults, with no overlap — which makes the physics/sensor distinction a measured property of the system rather than a labelling convention, and a defensible axis for the H3 generalization split.
Real headroom for an adaptation method to demonstrate an effect.
| Fault (severity) | Degraded | Vel. drop | Recovery | Falls | Recovery time |
|---|---|---|---|---|---|
| actuation_delay (10 steps) | 100.0% | 85.9% | 0.0% ± 0.0 | 0.0% ± 0.0 | — |
| torque_limit (0.2×) | 95.2% | 55.1% | 41.2% ± 3.0 | 33.2% ± 10.3 | 1.864 s ± 0.618 |
| joint_lock | 91.2% | 44.9% | 54.9% ± 18.6 | 0.0% ± 0.0 | 2.222 s ± 0.628 |
| sensor_dropout | 95.7% | 28.3% | 81.4% ± 14.4 | 0.0% ± 0.0 | 2.077 s ± 0.940 |
| sensor_noise (σ 0.3) | 83.2% | 23.7% | 86.8% ± 5.6 | 0.0% ± 0.0 | 2.227 s ± 0.451 |
Recovery is conditional on degradation — of the trials where the fault measurably slowed the robot, the fraction that returned to pre-fault speed and held it. Trials the fault never affected are counted under Degraded rather than scored as recoveries. torque_limit is the highest-power test case: moderate headroom, very low between-seed variance, and the only fault that topples the robot.
Training reliability
- Seeds attempted
- 10 @ 1.5M
- Converged
- 6 / 10 · 60%
- Excluded
- 1, 3, 5, 9
- Forward speed
- 0.585 ± 0.053 m/s
- Commanded
- 0.500 m/s
- Wall clock
- ~18.6 h
A 60% convergence rate is a reported result, not a defect to hide — it characterises how reliably this setup produces a usable policy.
Per-seed recovery range
- actuation_delay
- 0 – 0%
- torque_limit
- 39 – 45%
- joint_lock
- 24 – 76%
- sensor_dropout
- 55 – 93%
- sensor_noise
- 78 – 95%
joint_lock varies genuinely across seeds — some learned gaits tolerate a frozen joint, others don’t. Flagged automatically when the spread exceeds 40 points.
Build status
- Fault-injectable environment
- PPO base policy · 10 seeds
- Recovery criterion rebuilt
- Baseline A evaluation
- Across-seed aggregation
- Severity sweep · built, not run
- H3 held-out fault split
- Residual adaptation module
- Baseline B · full retraining
03 · Experience
Ten roles, running at once.
-
01
Active
Student Researcher
Massachusetts Institute of Technology
Member of MIT CSAIL, assisting development of a system that changes the way data can be visualized.
Jul 2026 — Present
-
02
Active
Student Researcher
University of California, Santa Cruz
Machine learning applications in robotics, and their integration into the real world.
Jul 2026 — Present
-
03
Active
Research Fellow
Lumiere Education
Conducting research alongside Fernanda-Maria Lugo-Bolanos from Brown University.
Jun 2026 — Present
-
04
Captain
Team Captain
FIRST Tech Challenge · Team 36541
Leads the team and teaches robotics fundamentals to rookie teammates.
Sep 2025 — Present
-
05
Active
Treasurer
VEX Robotics · Team 95330Z
Head of programming and CAD for the VEX robot, Team 95330Z.
Sep 2025 — Present
-
06
Founder
Founder
Apex STEM
Founded and runs a STEM initiative from the ground up.
Dec 2024 — Present
-
07
President
President
RIHS Math Club · Lathrop, CA
Leads the math club at River Islands High School.
Nov 2025 — Present
-
08
Active
Director
Binary Tree
Contributing to educational programming and outreach initiatives.
Jun 2026 — Present
-
09
Completed
Academic Writing Intern
TMAS Academy
Wrote a free AMC 8 prep book for accessible math learning — currently in the process of being published.
Jun 2025 — Aug 2026
-
10
Active
Troop Guide
Boy Scouts of America · Manteca, CA
Mentors younger scouts in outdoor skills and leadership.
Nov 2023 — Present
No entries in this discipline.
04 · Education
Two transcripts, one track.
River Islands High School
Sophomore year, 4.86 weighted GPA. Electives: AP Computer Science Principles, STEM Leadership, and AP Psychology.
- GPA
- 4.86 W
- Class of
- 2029
San Joaquin Delta College
Early College Pathway program, 4.0 unweighted GPA. Currently enrolled in CS11, POLS C1000, and PSYC C1000.
- GPA
- 4.0 UW
- Track
- Early College Pathway
05 · Recognition
Awards
MATHCOUNTS
Took first place at the MATHCOUNTS competition.
AIME Qualifier
Scored in the top ~14% of competitors on the AMC 12.
Stanford Math Tournament
Top 25% of competitors.
Berkeley Math Tournament
Top 20% of competitors.
USACO Silver
Silver division standing in the USA Computing Olympiad.
Congressional App Challenge
Built LifeLine, an app for people affected by natural disasters.
06 · Contact
Let's build something
Open to research collaborations, AI/ML internships, and math problems worth arguing about.