Robotics · AI/ML · Software

Jaiveer Bassi

I build and audit intelligent systems, from robot data pipelines and fault injection rigs to reproducible machine learning research.

I am a Robotics Software Engineer at a YC stealth startup and an AI Robot Operator at Physical Intelligence. I hold an M.S. in Software Engineering and a B.S. in Computer Science.

View résumé GitHub ResearchGate

Open to full time robotics software, AI/ML, and software engineering opportunities.

Current focus

Research

Reproducible work with the claim, evidence, controls, and limitations kept together.

Trajectory smoothness metrics under episode length controls

Accepted at NeurIPS 2026 RoboPADRobot learning · Demonstration curation

Smoothness Ranks Skill, Not Success: An Audit of Published Trajectory-Smoothness Curation Metrics Under Episode-Length Controls

Shows that published smoothness metrics for curating robomimic demonstrations mostly track episode length and operator skill rather than task success; curating by the contact-aware TED metric cut downstream policy success by a third versus random selection. Accepted as a poster at the NeurIPS 2026 RoboPAD workshop (arXiv:2610.17427).

Code & papers arXiv Base preprint WEBP OpenReview RoboPAD poster (OpenReview)Oops OpenReview

Closed loop success and open loop error across five demonstration failure modes

Under review at CoRL 2026 LfCRobot learning · Data quality

Not All Bad Demonstrations Are Equally Bad: Quantifying How Demonstration Failure Modes Degrade Closed-Loop Policy Performance

A controlled study of 660 policy fits across five failure modes, two policy families, ten seeds, and transition matched controls. Under review at the CoRL 2026 Learning from Corrections and Interventions workshop.

Code & results PDF OpenReview

Square task success rates by arm at 50 rollouts versus 500 rollouts per checkpoint

Under review at CoRL 2026 SimBench2RealRobot learning · Benchmark evaluation

What Can Three Seeds and Fifty Rollouts Resolve? A Preregistered Audit of Operator and Seed Variance on robomimic Multi-Human Benchmarks

A preregistered audit of 138 BC-RNN runs that splits robomimic Multi-Human benchmark variance into seed, demonstration, operator, and rollout noise. At 50 rollouts, ten seeds were statistically indistinguishable on every task; re-evaluating 46 Square checkpoints at 500 rollouts made the seed effect detectable and erased an apparent operator-proficiency pattern. Under review at the CoRL 2026 From Simulation Benchmarks to Real-World Robot Learning workshop.

Code & data PDF Preprint OpenReview Zenodo archive

Published HumanEval scores compared with the primary local pipeline and the no-newline diagnostic condition

Technical report v1.9Code models · Evaluation pipelines

One Byte, One Rank Reversal: Auditing HumanEval Pipeline Sensitivity for Base Code Models

A pinned five model reproduction on an RTX 4070 found StarCoder2 3B scoring 3 of 164 instead of its published 31.7 percent, reversing a published ranking. Removing one trailing newline from the evaluation prompt raised it to 49 of 164 and restored the order.

Code & report PDF Versioned archive

FP32 and INT8 latency distributions with backend specific speedups

Archived preprintEdge AI · Host CPU inference

When Does INT8 Actually Accelerate Host-CPU Inference? A Reproducible Audit of MobileNetV2 Conversion and Packaging

Shows that INT8 latency depends on backend delegation and quantization granularity, alongside paired accuracy and packaging controls. Version 1.2.0 is archived on Zenodo and has not been submitted for peer review.

Code & manuscript PDF Zenodo release

Build, break, measure

Selected projects

Robotics infrastructure, fault characterization, ML tooling, and production software. More on my GitHub.

Live π0 FAST shadow inference dashboard for the four degree of freedom teleoperation arm

π0 FAST · LoRA · RTX 4070

π0 on a Budget

Fine-tuning π0-FAST on a consumer RTX 4070 using synchronized demonstrations from a custom four-degree-of-freedom teleoperation arm, followed by open-loop and closed-loop evaluation. The live camera, Arduino telemetry, RTX policy server, and real-time shadow-inference dashboard are operational; real demonstration collection and policy fine-tuning are next.

GitHub →

CAD render of the Scribbot arm with its pen on a sticky note

OpenSCAD · Arduino · Servo control

Scribbot

A low-cost, mostly 3D-printed desktop robot arm that draws on a 3 × 3 inch sticky note: an MG90S base on a 2:1 printed gear, MG996R shoulder and elbow, an MG90S wrist, and a gravity pen holder, driven by an Arduino Nano and PCA9685 servo driver, with OpenSCAD sources, printable STLs, and stage-by-stage build guides.

GitHub →

Same simulated scene, different language instructions, different robot behavior

openpi · π0.5 · LIBERO

openpi Language Steerability Demo

Runs the public π0.5 LIBERO checkpoint in a single simulated scene with different language instructions and scores success with the simulator's own goal checks. Trained and paraphrased prompts succeed 10 of 10 times, while novel object recombinations drop as low as 1 of 10.

GitHub →

Simulated UR5e running a waypoint cycle with live joint angle telemetry

URSim · RTDE · Fault injection

UR5e RTDE Harness

Drives a simulated UR5e over RTDE, the same interface used by production UR5e cells, logs synchronized joint telemetry, and deliberately induces connection faults to characterize whether the client fails with a clean exception, a hang, or a native crash.

GitHub →

Closed-loop rollouts of policies trained at increasing demonstration contamination

Robot learning · Imitation · Simulation

Demonstration Quality Robustness

A controlled simulation study of 660 policy fits measuring how specific demonstration failure modes degrade closed-loop policy performance, and how poorly open-loop evaluation tracks that damage.

GitHub →Paper (PDF) →

Simulated two link arm grasp with bounding box, contact keypoint, and failure mode overlays

Computer vision · Labelbox · COCO

Grasp Annotation Pipeline

Samples public robot arm video into frames, provisions a Labelbox project with a grasp-focused ontology, and exports completed labels as COCO JSON. Each frame can carry a grasp_event bounding box, an object_contact keypoint, and a failure_mode label: drop, miss, collision, timeout, or success.

GitHub →

Low cost leader follower teleoperation rig

Arduino Nano and breadboard wiring for the teleoperation rig

Article · Teleoperation · Robotics

A $30 Teleoperation Rig

Rebuilding the core idea behind GELLO style robot demonstration collection with direct joint space control.

GitHub →Read on Medium →

USB to CAN interface used to monitor the bus

Article · CAN bus · Diagnostics

Simulating Real Robot Failures on a Breadboard

A three node CAN bus built to induce and diagnose failures that software logs alone cannot explain.

GitHub →Read on Medium →

Chess RL self-play games and training progress

C++ · PyTorch · Self play

Chess RL Engine

An AlphaZero inspired engine that learns only through self play: a C++ bitboard move generator exposed through pybind11, a PyTorch policy and value network on an 18 channel board encoding, and a PyGame interface showing move probabilities in real time.

GitHub →

Animated map of an MLP learning to classify shapes, showing activations, weights, and dropout

PyTorch · Matplotlib · Visualization

Neural Net Mapper

Trains an MLP on a synthetic shapes dataset and renders an animated map of its inner workings: neuron activations, weight signs and magnitudes, dropout, predictions, and live loss and accuracy curves.

GitHub →

Flask · PubMed · NLP

MedStract

Retrieves peer-reviewed abstracts from PubMed and turns them into summaries calibrated to four reading levels, from the general public to the domain expert, with question answering over the results, publication-trend charts, and citation exports in six formats.

Live site →