Caster Design Research¶

Researched July 2026. Context: the printed hemisphere caster (Caster.step) has zero tolerance for uneven floors — slight slopes strand the robot and rugs are a no-go (issue #3). The central constraint: any added compliance must not introduce wobble, because lidar scan-plane tilt degrades SLAM/localisation.

Problem Statement¶

Mote stands on two centred drive wheels plus caster support both fore and aft of the axle. That makes four ground contacts for a rigid body — statically over-constrained. On any floor that isn't perfectly coplanar with the build (and given FDM print tolerance, even a flat floor isn't safe), the chassis rocks fore/aft between the casters and the drive wheels unload → no traction → stuck. This is the observed failure; slopes and rug edges just make it worse.

Three things follow, and they frame every option below:

Caster shape cannot fix the stance. A better hemisphere, a bought nylon ball, a wheel — any rigid third/fourth contact leaves the statics unchanged. This matches testing: both printed and purchased ball casters were tried with no real improvement.
A 3-point stance is off the table. Turn-on-the-spot requires the centred axle, and the CoM straddles it: slightly behind for the base robot, moving in front when the SO-101 arm is fitted. Support is genuinely needed both fore and aft, and the CoM can't be shifted (the power bank is already crammed against the lidar between the servos).
The remaining degree of freedom is vertical compliance at the casters — a few mm of sprung travel so the wheels stay loaded and contact-height error stops mattering. The question this doc answers is whether that can be done without trading traction problems for lidar wobble.

Geometry and height budget¶

Quantity	Value	Source
Drive wheel	Ø65 mm (r = 32.5 mm)	`robot.yaml`
Caster positions	≈ ±100 mm from the axle	URDF (`caster_x`)
Ride height (base-plate underside → floor)	~16 mm	measured on robot
Current printed caster	Ø30 hemisphere, ~14 mm tall	`Caster.step`
Lidar scan plane	~54 mm above floor	URDF (`lidar_ground_height`)
Extra height without cutting the top plate	~5 mm	larger wheels, issue #3
Further gain with through-plate caster mounting	~6–7 mm	issue #3
Robot mass	~2 kg (estimate, more with arm)	BOM

So the caster envelope is ~16 mm as built, ~21 mm with bigger wheels, and ~27–28 mm with bigger wheels plus cutting mounting holes so a caster hangs from the top face of the base plate. Every bought "small" caster below is measured against those numbers.

What wobble costs¶

The lidar scans at only ~54 mm. A downward scan-plane tilt of 1° grounds the beam at ~3.1 m; 2° at ~1.5 m — floor returns appear as phantom walls right in the middle of useful range, corrupting both slam_toolbox and AMCL. Kinematic-ICP additionally assumes planar motion, so pitch transients leak directly into odometry error. Two distinct mechanisms matter:

Static/slow tilt — a sagging or unevenly compressed mount tilts the scan plane persistently. This is the dangerous one for mapping.
Transient rocking — the current rigid over-constrained stance already produces this: every accel/decel slams the chassis from one caster to the other, a hard-contact impact with the full contact-height error as amplitude. Rigid is not the wobble-free baseline; it is the current wobble source.

Option Survey¶

1. Optimised rigid printed caster (shallower angle / slider foot)¶

Reprofile the hemisphere: a flat furniture-slider style foot with a rounded, shallow (<30°) lead-in ramp instead of a curved face that presents >45° at rug height, optionally with a PTFE glide insert for low friction (PTFE furniture glides are a commodity — screw-on and adhesive discs in exactly this size range).

Uneven floors/rugs: fixes only the rug-edge attack angle; does nothing for the over-constraint, so slopes still strand the robot.
Wobble: none added.
Height: fits trivially.
Cost: filament (+ ~£4 for PTFE glides).

Verdict: necessary but not sufficient. The shallow-ramp, low-friction foot is the right tip geometry and carries into the recommendation — but as a rigid part it cannot fix the failure mode.

2. Bought ball casters / mini ball transfer units¶

Pololu-style ball casters (3/8″–1″ ball, ~10–18 mm heights) and industrial mini ball transfer units fit the height budget and roll omnidirectionally.

Uneven floors/rugs: already tested on Mote — no improvement, as the statics predict. Small balls additionally dig into carpet pile, and open bearing races ingest carpet fibre and hair until they seize (the classic small-robot complaint).
Wobble: none added (rigid), but inherits the rocking of any rigid stance.
Height: fits.
Cost: ~£3–8 each.

Verdict: rejected on evidence. Rigid contact in a nicer package.

3. Wheel casters (swivel)¶

The smallest commodity swivel casters start around 30–40 mm overall height — above even the stretched ~28 mm budget, so this path requires cutting the top plate or a bigger wheel redesign than the 5 mm allowance. And two new problems arrive with the swivel:

Caster flutter/shimmy at speed, and trail kickback: every direction reversal makes the swivel flip 180°, producing a lateral jerk right when the controller reverses — a yaw disturbance wheel odometry can't see.
Still rigid vertically, so the over-constraint remains.
Uneven floors/rugs: good over bumps a rolling wheel can climb; stance problem unsolved.
Wobble: adds yaw/lateral disturbances (flip, flutter).
Height: does not fit, even stretched.
Cost: ~£3–6 each, plus wheel/chassis rework to gain the height.

Verdict: rejected. Doesn't fit the budget, and the swivel dynamics are actively hostile to odometry on a robot this light.

4. Sprung drive wheels (the robot-vacuum approach)¶

Robot vacuums invert the problem: drive-wheel modules are spring-biased downward (iRobot's patents describe 5–25 N of bias per wheel), and the chassis rides on rigid front (and sometimes rear) casters. Wheels follow the floor; the body stays put.

Uneven floors/rugs: the gold standard — traction is guaranteed by spring preload over large travel.
Wobble: mixed. The chassis attitude is set by the rigid casters, so every bump under a caster still pitches the lidar directly; and traction is capped at spring preload rather than robot weight.
Height/complexity: a major redesign — the STS3215s are hard-mounted to Motor Support blocks between the plates; sprung modules need vertical travel, guides, and cable service loops in space that doesn't exist. It also decouples wheel odometry from body motion during suspension travel.
Cost: highest of any option (custom sprung modules).

Verdict: rejected for Mote. Right answer for a 3 kg vacuum designed around it from day one; wrong retrofit for this chassis, and it doesn't even remove caster-induced pitch.

5. Articulated rocker (linked fore/aft casters)¶

Join the front and rear casters with a beam pivoting on the chassis — four contacts become three effective ones (wheels + rocker), statically determinate, like a mini rocker-bogie.

Uneven floors/rugs: kinematically exact; handles anything within pivot range.
Wobble: the pivot is a genuinely free DoF, so the chassis pitch is determined — good — but the mechanism needs a stiff, low-slop pivot or the determinacy is lost to rattle.
Height/complexity: fatal — the beam must run fore–aft under the base plate through the space occupied by the power bank and lidar mount, inside a 16 mm envelope. There is no route.
Cost: printed, but the most moving parts of any option.

Verdict: rejected on packaging. The theoretically cleanest fix, worth remembering if the chassis is ever redesigned from scratch.

6. Compliant (preloaded plunger) casters — ⭐ recommended¶

Replace each rigid caster with a two-part printed unit: a casing bolted to the ORP grid on the base plate, and a plunger sliding vertically inside it with a rounded shallow-ramp tip (option 1's geometry) and ~4–5 mm of travel. A force element in the plunger's hollow Ø12 core reacts against the chassis plate itself (the plate is the cap — the only way the mechanism fits a 16 mm ride height, since the small tip retracts up through the casing lip rather than needing foot clearance below it). The casing lip retains the plunger; it inserts from the top before the casing is bolted on.

Force element candidates:

Element	Force curve	Notes
Compression spring (from a ~£6 assortment kit)	Linear, preload settable by fitted length	Best for wobble: firm, constant preload at ride height. Metal — no creep.
Opposing Ø12×3 N42 magnet pair (~£5 for spares-drawer quantities)	Steeply rising (~soft at rest gap, firm near contact)	"Standard printer parts", no sourcing precision. But the curve is backwards for this job: weakest exactly at ride height, where preload matters. Usable only if installed pre-compressed.
Printed TPU/PLA flexure	Printable, zero BOM	Creep/compression-set under constant load → ride height and preload drift. Rejected.
Off-the-shelf spring plunger (M8/M10)	Linear, ready-made	Tiny ball tip (Ø5–6 mm) digs into carpet; travel typically only ~2–3 mm. Rejected as the contact, could serve as an internal force element but the spring alone is simpler.

Recommendation within the option: compression spring, sized per the analysis below; keep the magnet pair as the documented no-spring alternative with the pre-compression caveat.

Uneven floors/rugs: the travel absorbs contact-height error, slope transitions, and rug pile up to ~4–5 mm while the preload keeps pressing the tip down and — critically — keeps the drive wheels loaded. Rug edges are handled by the shallow-ramp tip plus the traction that now exists.
Wobble: analysed in full below — net improvement over rigid.
Height: fits the existing 16 mm ride height. No bigger wheels, no top-plate cuts, no through-plate mounting, lidar height unchanged.
Cost: filament + ~£6 springs (or ~£5 magnets); optional PTFE tip insert ~£4.

Wobble vs Compliance Analysis¶

The trade-off the issue asks about, with numbers. Model: casters at L = ±100 mm from the axle, spring rate k per caster, robot mass ~2 kg.

Preload keeps the stance defined. With the CoM ~15 mm behind the axle, the casters must supply a net restoring moment of only 2 kg × 9.81 × 0.015 m ≈ 0.3 N·m, i.e. ~3 N more force at the rear caster than the front. Setting preload at ~3–4 N per caster at ride height covers this with margin while leaving ≈ 12–13 N (~65 % of weight) on the drive wheels — traction improves over today, where rocking can take wheel load to zero. Braking/acceleration pitch moments are tiny by comparison: at 0.5 m/s² and ~40 mm CoM height, the load shift at a caster is ~0.4 N ≪ preload, so contact never breaks and the chassis never crosses a free-play gap — the impact transient that defines the current rigid rocking simply has no mechanism.

Stiffness sets the pitch mode. With k ≈ 0.8 N/mm per caster, pitch stiffness is 2kL² ≈ 16 N·m/rad. Against a pitch inertia of ~0.016 kg·m² (2 kg at ~90 mm radius of gyration) the pitch natural frequency is ~5 Hz — well above drive-command content (< 2 Hz), so accel/decel excites the suspension quasi-statically rather than ringing it. The printed sliding fit adds friction damping for free. Softer springs push the mode down toward the excitation band and increase sag; much stiffer ones stop absorbing floor error. k ≈ 0.5–1 N/mm with 3–5 N preload is the design window.

Bounded, smooth tilt instead of impacts. Compliance does not eliminate pitch — a robot that tolerates a 4 mm floor step must pitch while crossing it (a 4 mm differential across the 200 mm caster span is ~1.1°). What changes is the character: the rigid stance takes that error as a hard impact with wheel unloading; the sprung stance takes most of it into the spring (the wheels, not the caster, are the stiff reference), leaving a smooth, friction-damped fraction-of-a-degree chassis motion. For the lidar the comparison is: occasional bounded ramps vs today's per-accel impact steps plus getting stuck. Static tilt is bounded by preload matching (front/rear springs fitted to the same length; a 1 mm asymmetry across 200 mm is < 0.3°) and by using metal springs so nothing creeps.

Why not magnets first: the repulsion curve means the preload at ride height is the weakest point of the stroke, exactly where the stance must be firm; the effective rate then rises steeply through the travel. The same parts work acceptably if the pair is installed with the gap pre-closed (preload built in), but a spring achieves the target curve without the workaround.

Recommendation¶

Fit preloaded plunger casters (option 6) at both the front and rear positions: printed casing + printed plunger, rounded shallow-ramp tip, compression spring in the hollow core reacting against the base plate, ~4–5 mm travel, ~3–4 N preload, k ≈ 0.5–1 N/mm. Combine with option 1's tip geometry (shallow lead-in, optional PTFE insert).

Rationale: it is the only option that (a) addresses the actual failure — the over-constrained rigid stance — rather than the caster's packaging, (b) fits inside the existing 16 mm ride height with no wheel, plate, or lidar-height changes, (c) reduces lidar disturbance relative to today by replacing hard rocking impacts with preloaded, friction-damped, bounded compliance, and (d) costs a few pounds of commodity hardware.

BOM delta (vs `BOM.md`)¶

Change	Part	Qty	Unit price	Notes
−	Caster (3D printed hemisphere)	1	(filament)	removed
+	Caster casing + plunger (3D printed, PLA)	2 sets	(filament)	front and rear
+	Compression spring assortment kit	1	~£6	Amazon UK; pick ~Ø8–10 mm, k ≈ 0.5–1 N/mm
+ (alt.)	Ø12×3 mm N42 disc magnets (opposing pair per caster)	4	~£5/pack	first4magnets; only with built-in pre-compression
+ (opt.)	PTFE glide insert for plunger tip	2	~£4/pack	commodity furniture glides

Net delta: ~£6–10.

Implied CAD / mounting changes (follow-up task, not this one)¶

Retire Caster.step; author a two-part casing (bolts to existing ORP grid holes on the base-plate underside — no new holes in either plate) and plunger (rounded shallow-ramp tip, hollow Ø12 core, lip-caught body, inserted from the top). The base plate itself is the spring's upper seat.
Two units: the front position and a rear position on the same grid.
Update ASSEMBLY.md (print table, step 3, drop the "provisional" caveat) and BOM.md per the delta.
Update the URDF caster properties (caster_radius, caster_x, caster_z are currently marked assumed) to the as-built values, and add the rear caster link.
Fit check on hardware: confirm ride-height preload front/rear, then a mapping run over a rug edge as the acceptance test — wheels must stay loaded and the map must stay clean.