Research
The research group studies the hardware and software problems that recur across teams in the network — localization, perception, and match-scoring analysis among them — and publishes its methods and results so any team can evaluate or reuse the work. Entries below list the group's formal write-ups; the updates that follow report on work still in progress.
Research entries
Latency Characterization of the Shared Vision Pipeline
June 1, 2026We measure end-to-end latency, from image capture to bounding-box delivery over serial, for the shared game-piece detection pipeline running on two candidate single-board computers under match-representative load. The lower-cost board under evaluation added a median 46 ms of latency relative to the board currently used in pilot deployments, a difference we judge acceptable for the majority of control loops surveyed but that narrows the safety margin for the fastest-cycling mechanisms in the network.
Synopsis, methods, and results
Testing whether a cheaper onboard computer can run the shared vision pipeline without slowing it down enough to matter. It runs about a twentieth of a second slower than the current board, which is fine for most robots but worth knowing about before recommending it network-wide.
Motivation
The vision pipeline’s midseason progress report identified board selection as the remaining open question before a wider release: the single-board computer used in the three-team pilot performs well but adds a nontrivial cost to a team’s build. A lower-cost board was proposed as an alternative, but its detection latency under realistic load had not been benchmarked. This study measures that latency directly rather than relying on the vendor’s general-purpose specifications, which are not representative of a sustained detection workload competing with the board’s other processes.
Method
We instrumented both candidate boards to timestamp three points in the pipeline: frame capture from the camera sensor, completion of the detection model’s forward pass, and transmission of the resulting bounding-box coordinates over the serial link to a simulated downstream controller. Each board ran the identical detection model and camera at a fixed frame rate, under a synthetic background load intended to approximate a robot’s other onboard processes (telemetry logging and a simulated control loop at 100 Hz). We collected 4,000 detection cycles per board across two capture sessions on different days to account for thermal throttling effects, which we observed on the lower-cost board during sustained runs.
Results
The current pilot board produced a median end-to-end latency of 61 ms, with a 95th-percentile latency of 88 ms. The lower-cost candidate produced a median latency of 107 ms, with a 95th-percentile latency of 164 ms under sustained thermal load — the gap widens under load because the candidate board throttles its clock speed after roughly four minutes of continuous inference, an effect not visible in short benchmark runs. For the majority of mechanisms surveyed across network teams, control loop cycle times are long enough that either board’s latency is a small fraction of the loop period. A small number of high-speed intake and indexing mechanisms cycle fast enough that the additional 46 ms median latency would consume a meaningful share of their control budget.
Status and Next Steps
These results are preliminary: we have not yet tested either board’s behavior with active cooling, which the lower-cost board’s manufacturer suggests would substantially reduce throttling, nor have we tested detection accuracy under the same thermal conditions, which may also degrade. A full manuscript with the complete data set, thermal profiles, and a board recommendation is in preparation and will be made available once active-cooling testing is complete, expected within the current term.
References
- Vision pipeline midseason progress report, published to the community update feed, 2026-05-19.
- Manufacturer thermal design specifications for both evaluated boards, retained with the raw test data.
Encoder-Based Odometry Calibration Across Chassis Variants
March 10, 2026We characterize accumulated position error from wheel encoder odometry on three chassis configurations in common use across the network and evaluate a per-unit calibration procedure that corrects for wheel diameter tolerance and encoder mounting offset. Calibration reduced median drift over a two-minute run by 34 percent relative to manufacturer nominal values, with the largest gains observed on six-wheel drivetrains where uncorrected wheel diameter mismatch between the two sides of the chassis was the dominant error source.
Synopsis, methods, and results
A short calibration routine that measures each robot's actual wheel size and encoder alignment cuts navigation drift by roughly a third, with the biggest improvement on six-wheel chassis where small manufacturing differences between wheels add up over a match.
Motivation
Dead-reckoning odometry from wheel encoders remains the most common localization method among network teams because it requires no additional hardware beyond what most control systems already include. Its accuracy, however, depends on constants — wheel diameter, encoder counts per revolution, and track width — that are typically taken from a part’s nominal specification rather than measured on the assembled robot. Manufacturing tolerance on commonly used wheels runs as high as 2 percent of nominal diameter, and encoder mounting introduces a further small but consistent offset. Both errors are systematic rather than random, so they accumulate rather than average out over a run.
Method
We tested three chassis configurations drawn from robots currently fielded by network teams: a two-wheel differential drive, a four-wheel skid-steer, and the six-wheel configuration used in the second-generation chassis platform. For each configuration we ran five physical units through a fixed 4 meter square course with painted reference lines, recording encoder counts and comparing the odometry-estimated final position against the measured true position.
The calibration procedure itself is a one-time routine performed on each assembled robot: it drives a short fixed pattern (forward, rotate in place, forward), compares encoder-reported displacement against the known course geometry, and solves for effective wheel diameter and track width that minimize the discrepancy. The routine takes under two minutes and requires no equipment beyond a flat floor with two known reference points.
Results
Prior to calibration, median accumulated position error after a two-minute run across all fifteen units was 18.4 cm. After calibration, median error fell to 12.1 cm, a 34 percent reduction. The improvement was not uniform across configurations: two-wheel differential drive units, which have only one wheel diameter parameter per side to correct, improved by 21 percent, while six-wheel units improved by 46 percent. We attribute this difference to the six-wheel platform’s greater sensitivity to left-right wheel diameter mismatch, since uncorrected error there compounds across three wheel pairs per side rather than one.
Calibration constants were stable when re-measured after one week of normal use, with encoder count and wheel diameter estimates changing by less than 1 percent, suggesting the routine does not need to be repeated often under normal wear.
Limitations and Next Steps
All testing was performed on a smooth indoor floor with the robot’s stock wheels; drift under partial wheel slip, such as on textured competition mats, was not evaluated here and is the subject of ongoing work referenced in the group’s midseason progress notes. We also did not test calibration stability after a wheel replacement, which teams do routinely after damage, and would expect to require re-running the routine.
References
- Borenstein, J., Feng, L. “Measurement and Correction of Systematic Odometry Errors in Mobile Robots.” IEEE Transactions on Robotics and Automation, 1996.
- Internal test log and calibration firmware, available on request alongside the manuscript.
Request the manuscript
Research updates
- 2026