Reliable robot perception.
LiDAR and machine learning.
My research aims to improve how robots detect obstacles from sensor data: fewer missed objects, controlled false alarms, and timely detection. I work with LiDAR and investigate camera evidence where it can help.
How do robots turn scattered measurements into reliable evidence of an obstacle?
Explore one of my recorded Mid-360 scans. Rotate the scene to see how individual measurements describe its structure.
The interactive scan could not load. Choose “The reference” to see the original research figure.
Same recorded points in two coordinate frames.
Original analysis · 13 October 2026
Recorded scan · 14,277 filtered points. Color shows relative height; this view is not a detection result.
I explored hyperdimensional spatial memory and reliability signals for LiDAR-inertial odometry. The work compared spatial representations under equal memory budgets and examined whether motion-estimation diagnostics provided useful warnings.
Outcome. The tested spatial-memory approaches did not establish a utility advantage over explicit records at matched storage budgets. For AKF-LIO and LOG-LIO2, constant uncertainty scales fitted on one development trajectory missed the preregistered coverage band on all 12 evaluation trajectories. These routes are closed; they informed the move toward directly measurable obstacle-perception problems.
I reproduced released NDP and LIDO models on STU validation and examined score thresholds, clustering, ground removal, and false objects per scan. The study asks what an object-level benchmark score actually measures.
Finding. Removing points close to the ground made NDP’s clusters fit object boundaries better. At NDP’s selected thresholds for the one-false-object-per-scan budget, the objects credited by the improvement were already flagged. Higher overlap-based recall did not establish that more obstacles had been found.
Validation-only analysis: 19 sequences and 30 sequence–instance pairs, with physical identity across sequences unverified. A measurement study, not a new detector or a published result.
Using Mid-360 recordings with a small box present and removed, I examined ground filtering and the sparse returns available close to the floor. The experiments compared short observation windows with accumulated measurements.
Finding. Accumulation increased detections in the selected region across four paired recordings of a 2.9 cm box. The result demonstrates the importance of observation time in this setup; it does not establish a new detection method.
One development setup; the box’s location was inferred from scene changes rather than independently measured. A subsequent mapping construction was stopped after its removal control exposed stale obstacle marks.
A bidirectional study of translation quality, localized errors, and human review.
I compared six translation systems across nine languages and English, evaluating both English → target language and target language → English. The study examined where automated estimates and localized error spans could support review of biblical text.
My contribution covered data collection, evaluation infrastructure, analysis, visualizations, and the working paper. The applied context was Trans World Radio (TWR).
Translation direction and language coverage affect the comparison. Aggregate scores also conceal local changes in meaning—an important consideration when deciding what people need to review.
Critical spans as a share of all predicted error spans, pooled across systems and both directions. This is not a translation-level error rate. System coverage differs; Yoruba is outside documented evaluator coverage.
These examples show why a quality score needs to be examined alongside the source text. Select a case to compare the wording.
When she had weaned Lo-ruhamah, she conceived and gave birth to a son.
When she had devoured Lo-ruhamah, she conceived and gave birth to a son.
“Divezzato” means weaned in this context. The output substitutes “devoured,” changing the action described.
Quando ebbe divezzato Lo-ruhamah, ella concepí e partorí un figlio.
“she had devoured Lo-”
Word reference · Treccani ↗The red word above is a selected word within this model-flagged passage.The evaluator identifies a passage for review. A qualified reader must assess the wording in context before the translation is used.
Examples are selected Amazon Translate outputs into English. They illustrate possible meaning changes, not comparative error rates. The English glosses are editorial readings; a critical flag does not establish unethical intent.
Choose a language and reverse the direction. Select a system to inspect its translation and the passages flagged by xCOMET.
Automated estimates, not accuracy percentages. Red: critical · Amber: major. Minor spans are hidden. Model flags require human review.
3D pose estimation and Parkinson’s-related movement research.
I worked on reconstructing 3D human pose and classifying gait from joint sequences. These were separate experimental paths: a PoseMamba reconstruction benchmark and a MediaPipe-based gait classifier.
The engineering included PyTorch training, CUDA compatibility, checkpoint handling, and MLflow tracking on DGX Spark.
I adapted and trained the published PoseMamba architecture to reconstruct 3D coordinates from 2D joints.
Best mean per-joint position error
Human3.6M · epoch 89 / 120
Final epoch: 43.63 mm. This measures pose reconstruction, not Parkinson’s diagnosis.
On 302 development clips, bone refinement and smoothing reduced macro F1 relative to the MediaPipe baseline.
| Processing | Macro F1 |
|---|---|
| MediaPipe baseline | 0.767 |
| Bone refinement | 0.723 |
| Bone refinement + smoothing | 0.733 |
These clips do not establish independent patient-level generalization.
Original ST-GCN chart using SMPL-X-derived joints, separate from the MediaPipe baseline. Three-fold snapshot; five folds later completed. Subject independence is unverified. These curves do not establish diagnostic performance.
Six systems · Nine target languages plus English · Both directions
I collected and aligned verse-level translations, normalized system exports with Python and pandas, ran GPU-based evaluation, analyzed the outputs, and prepared the figures and working paper.
xCOMET-XL provides segment-level quality estimates and localized error spans. COMET-Kiwi provides reference-free estimates; BLEU and chrF add lexical comparisons.
In the saved study, pooled Arabic xCOMET estimates averaged 0.714 into Arabic and 0.434 into English. Rankings changed when comparisons used common language coverage. Localized spans identified passages requiring closer review, including altered word senses and metaphor distortions.
The corpus contains 919 verse IDs from chapter one of 61 books. It is not a representative full-Bible sample. The 95 observed system–language–direction combinations have unequal coverage.
Scores are model estimates, not accuracy percentages or theological judgments. Yoruba is outside the evaluators’ documented language coverage. Historical system versions, prompts, run dates, and the reference-scoring configuration are incompletely documented; these results do not rank current products. Recovered COMET-Kiwi output covers 86,371 records.
Evaluating Machine Translation of Biblical Verses: A Bidirectional Multi-System Study with XCOMET-XL Segment Scores and Fine-Grained Error Spans.
University of Northwestern–St. Paul. Faculty mentors: Dr. Jonathan Zderad and Prof. Mark Antiel. Applied context: Trans World Radio.
xCOMET-XL model card ↗COMET-Kiwi XXL model card ↗
I adapted and trained the published architecture on Human3.6M. The 120-epoch log records a best MPJPE of 41.96 mm at epoch 89 and a final MPJPE of 43.63 mm. These are reconstruction results; PoseMamba was not part of the saved MediaPipe classifier’s inference path.
The baseline maps MediaPipe world landmarks to 17 joints and feeds sequences to an ST-GCN classifier. In a 302-clip development comparison, macro F1 was 0.767 for the baseline, 0.723 after bone refinement, and 0.733 with refinement and smoothing. The evaluation does not establish independent patient-level generalization.
The original chart shown on this site is from a separate ST-GCN experiment using SMPL-X-derived joints. Its early snapshot contains three folds; the saved log later completed five. Historical subject-independent splitting is unverified, and a later split correction exists. The chart does not establish clinical diagnostic performance.
PyTorch, CUDA compatibility work, DGX Spark training, checkpoint management, MLflow tracking, and skeleton-processing comparisons.
Improve reliable robot obstacle perception using LiDAR and camera evidence where justified. A useful contribution must reduce missed obstacles at comparable false-alarm rates and detection delays, and work beyond the examples used to develop it.
The offline Mid-360 pipeline preserves XYZ, intensity, tag, line, timestamps, and point identity before selection. A fitted floor frame provides a provisional height reference for the tilted mount. Same-recording consistency is not independent calibration; robot-base and dual-sensor calibration remain unresolved.
The cable study uses published segmentation weights and a compatibility-patched CVF-DLO tracing component on 12 synthetic development images containing 24 annotated cable instances. Those instances are not independent physical objects. Simple changes did not produce a robust improvement under the nominal false-pixel criterion. Real-world generalization and LiDAR benefit are unverified.
A proposed measured-cord experiment would separate absent raw returns from evidence lost downstream. It requires independently measured target geometry and has not been performed. Thin obstacles remain a feasibility candidate within the broader research goal.
The earlier spatial-memory and odometry routes are closed. The STU evaluation study is retained as a private working manuscript; its analysis uses validation data, and hidden-test scores are unavailable. The low-obstacle recordings are exploratory; the failed mapping construction remains closed.
The floor-frame figure is the original 13 October analysis of saved Mid-360 points, shown without alteration. The entrance uses 14,277 filtered points from an earlier recorded scan. Its colors and camera movement are presentation styling, not measured classifications or a live sensor feed.
I work on machine learning systems for language, human movement, and robot perception.
At the University of Northwestern–St. Paul, I studied biblical translation across six systems and nine languages in both directions, with Trans World Radio as the applied context. I built the collection and evaluation pipeline, analyzed the results, and prepared the research figures and working paper.
My subsequent work explored 3D pose reconstruction and Parkinson’s-related gait classification, including GPU training and experiment engineering on DGX Spark.
My research interests center on LiDAR and reliable robot perception: how sparse measurements, unfamiliar objects, and evidence across observations can inform what a robot notices.