Simulation and live trial, 1,260 pick runs, one warehouse

Robotics research poster example

A robotics poster has to show the method as a sequence before it shows a result, because the result only means something once the reader knows what the robot does at each step. This board runs map, plan, detect, replan and execute across the top as a five stage chain, then gives run times at three obstacle densities underneath.

Create a poster with OneCraftA0 Portrait, printed at 841 by 1189 millimetres

The whole board

The poster at full size, exactly as it prints. Every number, citation and caption on it was written for this example, so the layout is being judged on real content.

RU
Local replanning for warehouse pick runs: 1,200 simulated runs and 60 on the floor
A. Krisnadi ¹, L. Fenwick ¹, T. Mbeki ²
1 Autonomous Systems Laboratory, Ravensmoor University · 2 Industrial Robotics Centre, Ravensmoor University
Abstract
A pick robot that replans its whole route every time a person steps into an aisle spends more time thinking than driving. This work tests a local replanner that keeps the global route and resolves only a six metre window around the obstacle. It was run 1,200 times in simulation and 60 times on a working floor, at three obstacle densities, against the full replanning baseline the site uses today. Mean run time fell at every density, from 84 to 79 seconds at low density, 97 to 86 at medium and 121 to 98 at high, which is 19% faster where the aisles are busiest. Collisions, counted as any contact or emergency stop, fell from 28 to 10 across 630 runs of each method. The gain grows with density, which is the opposite of what the site expected and the reason the result is worth a board.
HOW THE STUDY RAN
1
Map
A static map of 4,200 square metres and 38 aisles, loaded once at the start of a shift.
2
Plan
One global route to the pick face, computed once and kept unless an event forces a change.
3
Detect
Lidar and a depth camera flag anything entering the lane inside 2.5 metres of the robot.
4
Replan
Only a six metre window around the obstacle is resolved, in a median of 38 milliseconds.
5
Execute
Drive the local window, then hand control back to the untouched global route.
Baseline
Proposed
Low
Mid
High
03570105140
Figure 1. Mean run time at low, mid and high obstacle density, local replanning against the baseline. 210 runs of each method at each density.
Source: 1,200 simulated and 60 live pick runs, 2026.
Results
Mean run time for the baseline was 84, 97 and 121 seconds at low, medium and high obstacle density. For the local replanner it was 79, 86 and 98 seconds. That is 6.0%, 11.3% and 19.0% faster, so the advantage roughly triples between the quietest and the busiest condition. Collision counts, where a collision is any contact or triggered emergency stop, run: low density 2 against 1, medium 7 against 3, high 19 against 6, out of 210 runs of each method at each density. Across all 630 runs of each method that is 28 against 10. The live runs behaved like the simulation. Sixty runs, 10 of each method at each density, gave 82, 89 and 103 seconds for the local replanner against 87, 101 and 126 for the baseline, every one of them inside 6% of the simulated mean for the same cell. Planning time accounts for most of the difference. The baseline solved a full route in a median of 310 milliseconds and did it twice, five times and eleven times per run at the three densities. The local replanner solved its window in a median of 38 milliseconds and was called the same number of times, so the thinking cost across a high density run fell from 3.4 seconds to 0.4.
REFERENCES
1.Halloway, P. and Ostrovsky, D. (2025) Windowed replanning against global replanning in cluttered aisles. Journal of Field Robotics Practice, 12(3), 188 to 207.
2.Nagy, B. (2024) Planning cost per event and why density changes the answer. Autonomous Systems Letters, 8(1), 41 to 56.
3.Mbeki, T., Krisnadi, A. and Fenwick, L. (2026) Sixty live runs: what simulation got right on a working floor. Ravensmoor Robotics Reports, 5(2), 63 to 80.
4.Ravensmoor Industrial Robotics Centre (2026) Obstacle density protocol for warehouse pick trials, version 2.
Discussion
The reason the gain grows with density is that the baseline pays its planning cost per event, and events arrive faster in a busy aisle. At low density the robot replans twice in a run and the difference is almost entirely drive time. At high density it replans eleven times, and each of those is a full route solve on a map of 38 aisles rather than a window of six metres. The collision result is the one to treat carefully. Nineteen against six is a large relative difference on small counts, and the intervals overlap at low and medium density. What can be said is that no collision in either method came from the local window failing to find a path; all 38 came from the robot being already inside the stopping distance when the obstacle appeared, which is a sensing problem rather than a planning one. Three things were not tested: a moving obstacle that tracks the robot, a floor with more than one robot replanning at once, and any map that changes during a shift. Each of those is a plausible way for the gain to shrink.
Key finding
Local replanning was 19% faster at high obstacle density, 98 seconds against 121, and cut collisions from 19 to 6 in that condition.
Run time 19.0% lower at high density, 98 s against 121 s
Gain grows with density: 6.0%, 11.3% then 19.0%
Collisions 10 against 28 across 630 runs of each method
Live runs within 6% of the simulated mean in every cell
Rig and runs
A single differential drive pick robot with a 2D lidar, a forward depth camera and a 120 kg payload. Simulation used the site map at 1:1 with 1,200 runs, 400 at each of three obstacle densities, split evenly between the two methods. Live testing added 60 runs on the working floor outside picking hours, 10 of each method at each density. Densities were set at 1, 3 and 6 obstacles in the aisle at once.

Block by block

What each block on the board is for, in the order a reader walks it.

Title band and three authors
The title names the method, the simulated run count and the live run count, so a reader knows immediately that this is not a simulation only result.
Abstract
Why replanning a whole route on every obstacle is expensive, what the local replanner does instead, the three density results, and the collision counts of 28 against 10 across 630 runs of each method.
Method as a five stage chain
Map, plan, detect, replan and execute across the full width, each with its own detail line. A sixth stage would be silently dropped by this block, so the method was written to five.
Figure 1, run time by density
Mean run time for both methods at low, medium and high obstacle density, 210 runs of each method behind each bar, with short tick labels and two short series names.
Results
The six means, the three percentage gains, the collision counts per density, the live run comparison, and the planning time that explains the difference: 310 milliseconds per full solve against 38 per window.
Discussion
Why the gain grows with density, why the collision result should be read carefully on small counts, what all 38 collisions had in common, and three things that were not tested.
Key message and four checked points
A callout carrying the 19% figure at high density, then a four item list giving the run time, the growing gain, the collision totals and the live run agreement.
Rig and runs, then references
The test setup section carries the robot, the sensors, the run counts and the three obstacle densities, because this design has no separate protocol block. Four numbered references close the board.

What makes this board work

Five stages, each with its number

The process chain does not just name the stages, it carries the detail that matters at each one: a 4,200 square metre map, a 2.5 metre detection distance, a six metre replanning window and a 38 millisecond median solve. A chain of bare nouns would explain nothing.

The gain is shown growing with density

Plotting three densities rather than one average is the whole point of the figure. It shows 6.0%, 11.3% and 19.0%, which is the finding: the advantage triples between a quiet aisle and a busy one, because the baseline pays its planning cost per event.

Live runs are reported beside the simulation

Sixty runs on a working floor landed within 6% of the simulated mean in every cell, and the board says so with the numbers. A simulation only poster invites the obvious question, and answering it in the results block costs three sentences.

Questions people ask

What goes on a robotics research poster?

The method as a sequence, the test conditions and how many runs at each, the baseline you are beating, the results with their spread, and any failure cases. If the work includes both simulation and hardware, report them separately so a reader can see how far the simulation carried.

How many stages fit in a method chain?

Five on this design, and a sixth is silently dropped rather than flagged, so write the method to the number the block declares. Only two of the forty poster designs carry a process chain at all, so if your method is a sequence, pick the design first and shape the stages to it.

Should a robotics poster show simulation or hardware?

Both if you have both, and label which is which. This board reports 1,200 simulated runs and 60 live runs and gives the live means beside the simulated ones, within 6% in every cell. That comparison is more persuasive than either result alone and it costs one paragraph.

This design has no protocol block. What then?

Move the content into a section the design does have. Here the rig, the sensors, the run counts and the obstacle densities went into the final section, retitled for the purpose, because the section headings on this block are editable while the headings on the method and reference blocks are fixed by the design.

How do I keep three categories readable in a figure?

Keep the tick labels to one short word and the series names short, because the legend sits in the bottom margin beside the tick row. This figure uses low, mid and high with two series named baseline and proposed, and spells the densities out in the caption instead.

What does an A0 portrait poster export at?

The working canvas is 1,656 by 2,340 pixels and the export captures at twice that, so 3,312 by 4,680 pixels, about 100 dots per inch at full A0 size. It is RGB with no bleed setting, no crop marks and no CMYK conversion, so a print shop that needs them will set them up.

Build your own in about a minute

The button below opens the generator with this use case already described. Change the wording to match your own, generate, then edit anything you like.

Make one like this

Other poster examples

Cybersecurity research poster example

A phishing study poster is usually read from its numbers down, so this board opens with four of them across the top: 3,200 staff, two campaigns, a click rate that fell from 18.4% to 9.1%, and a report rate that rose from 12% to 31%. Everything under the strip explains how those four were produced.

Electrical engineering research poster example

A bench test poster is only useful if the reader can see the whole curve, not the peak. This board runs one 5 kW inverter across five load points on three firmware versions, plots conversion loss rather than efficiency so the differences are visible, and shows that the same 0.9 point gain holds at 30% of rated load, where a rooftop array actually sits.

Renewable energy research poster example

A renewable energy research poster earns its place when it tests a number people rely on. This Australian A0 landscape board checks the design model behind rooftop solar quotes against a year of meter data from 220 systems, and finds the model overstates yield by 3.1% overall and by 7% on south facing roofs.

Want the steps in the builder? Read Add charts and diagrams, then choose the template, theme and size. For everything this generator can do, see the poster maker.

Sources

Written and checked by the OneCraft team. Last checked .