THE NEXT ROBOT HAS TO EARN EVERY NUMBER.
The humanoid work is paused. ATH is investigating a racket-bearing robotic arm on a driven wheeled base. The revised design target is about 59 m/s racket-tip speed (Daniel said 59.3 m/s is good enough; earlier studies screened 60 m/s), alongside broad on-court shot coverage, fast serves and high-spin groundstrokes. Daniel wants lever length optimized for tip speed without excess inertia or self-interference, alongside low distal moving mass, structural durability and controllable ball placement/spin across shots; those are design gates, not tested capabilities. That speed is a requested requirement, not measured ATH performance. The hardware, its physics model and those abilities are not validated.
ESTHER, not ATH
A published tennis robot mounted a 7-degree-of-freedom, cable-driven Barrett WAM arm and racket on a motorized tennis wheelchair. Cable drive is not the same as the QDD architecture under consideration. The researchers measured 10 m/s pre-impact racket-head speed and 1.42 m/s² autonomous base acceleration. In ball-machine court trials, 66% and 53% of shots were returned successfully in the two reported setups (15 trials each). The 60 m/s ATH design target is six times ESTHER's measured 10 m/s racket-head speed. This ratio compares targets, not proven capabilities. Those results belong to ESTHER, not our proposed robot. Read the paper.
Arm + driven base
Candidate layout: a powered wheeled chassis, a multi-joint racket arm, physical racket and stringbed contact, ball tracking and an inverse-kinematics planner. No arm, motor, gear ratio, wheel, power system or sensor stack has been selected as a verified ATH build. A component's catalog limit does not prove an assembled robot can perform a shot.
60 m/s drivetrain check: candidate chain fails
Verdict for the tested off-the-shelf QDD stack: no-go. Even an optimistic planar lower-bound model needs joint torques above all three selected peak catalog values to reach a 60 m/s racket tip. A custom drivetrain and a verified, full-chain dynamics solve are needed before ATH can claim this target. This does not prove that every off-the-shelf actuator arrangement or every custom design is impossible; it rejects only the documented candidate under the stated test.
Reference geometry, not a selected ATH design: the Barrett WAM 7-DOF DH table gives a 0.55 m upper arm and 0.30 m forearm; the Wilson Blade 98 listing gives a 0.685 m / 323 g strung racket. For a straight 3R planar swing, the shoulder, elbow and wrist lever arms to the far tip are approximately 1.535, 0.985 and 0.685 m. The 27-inch racket is below the ITF 73.7 cm maximum. Actual base placement, shoulder height, joint reach for serve contact, spin axis and court coverage are not yet defined and must be checked in 3D IK, not inferred from this slice.
48 V catalog stack: AK10-9 V2 KV100 shoulder - 960 g, 9:1, 421 rpm rated, 18 Nm rated / 48 Nm peak; AK80-6 KV100 elbow - 485 g, 6:1, 603 rpm rated, 6/12 Nm; AK60-6 V3 KV80 wrist - 380 g, 6:1, 490 rpm rated, 3/9 Nm. The first two vendor tables label inertia without specifying its reference frame; we have not reflected those figures through gearing. The AK60 table calls its 243.5 g cm² value rotor inertia, but even this value is excluded from the optimistic lower bound.
| 0.15 s rest-to-contact test | Shoulder | Elbow | Wrist |
|---|---|---|---|
| Maximum assumed joint sweep | 1.5 rad | 1.5 rad | 2.0 rad |
| Optimized contact speed | 137 rpm | 191 rpm | 255 rpm |
| Optimistic torque needed, minimum | 150 Nm | 66 Nm | 27 Nm |
| Vendor peak torque | 48 Nm | 12 Nm | 9 Nm |
| Needed / peak | 3.1x | 5.5x | 3.0x |
Method: constant nonnegative joint acceleration from rest, straight planar Jacobian v_tip = [1.535, 0.985, 0.685] dot q̇. The inertial mass matrix sums the 0.323 kg racket as a point mass at 1.18 m (its real distributed inertia is not yet measured), a 0.380 kg wrist module at 0.85 m, and a 0.485 kg elbow module at 0.55 m. These placements and the 1.5/1.5/2.0 rad sweep caps are test assumptions, not measured robot properties. A linear program minimizes the largest fraction of catalog peak torque, constraining each joint to its catalog rated speed and the assumed sweep. With all peak torques enforced instead, the mathematical tip-speed ceiling is 18.3 m/s in 0.15 s, 20.1 m/s in 0.20 s; these are upper bounds under this assumption set, not predicted ATH speeds. Missing links, fixture, moving cable/gear inertia, impact, compliance and loaded torque-speed/thermal curves could worsen the result. Do not mix a catalog peak torque with a rated-speed point and call it an attainable operating point.
Changing the swing time, joint ranges, distribution of racket mass or base motion changes the bound, but it cannot turn this simplified test into proof of a 60 m/s build. Next design gate: source a complete drivetrain and mechanically measured arm/racket inertias; solve 3D IK and inverse dynamics through contact, then verify coupled torque-speed, energy, power, current, bus sag, thermal duty cycle, base stability and measured racket/ball impact.
Variant: add a rotating torso
It helps, but this published four-axis stack does not reach 60 m/s in the stated test. A powered torso can add racket-tip velocity and permit a longer kinetic-chain wind-up. It also must transmit and accelerate all distal mass. Moving motors proximally, as in cable or remote actuation, may lower the moving inertia, but it does not itself prove any 60 m/s shot.
Scenario geometry: the preceding Barrett-manual 0.55 m upper and 0.30 m forearm, the strung Blade 98 0.685 m / 0.323 kg racket, and an assumed, not measured ATH 0.25 m horizontal offset from trunk yaw to shoulder. The Blade listing gives a 0.330 m strung balance point from its butt. With the straight planar axes at 0/.25/.80/1.10 m, the far-tip lever arms are 1.785/1.535/.985/.685 m for torso/shoulder/elbow/wrist. This geometry does not establish serve contact height or full-court reach; neither the mounting height nor the base envelope is selected.
Tested 48 V stack: CubeMars AKE90-8 KV35 torso (1.4 kg, 8:1, 55 Nm rated / 170 Nm peak, 120 rpm rated / 210 rpm no-load, external driver needed), with the same AK10-9 V2 KV100 shoulder (48 Nm peak), AK80-6 KV100 elbow (12 Nm peak) and AK60-6 V3 KV80 wrist (9 Nm peak). This is a comparison, not a hardware selection. The torso's 1.4 kg motor is assumed mounted at the trunk axis and stationary with respect to that yaw; a different mount changes its inertial load.
| Rest-to-contact test | Torso | Shoulder | Elbow | Wrist |
|---|---|---|---|---|
| Assumed maximum sweep | 1.5 rad | 1.5 rad | 1.5 rad | 2.0 rad |
| Catalog peak torque | 170 Nm | 48 Nm | 12 Nm | 9 Nm |
| Minimum torque to reach 60 m/s in 0.15 s | 189 Nm | 151 Nm | 66 Nm | 27 Nm |
| Minimum torque to reach 60 m/s in 0.20 s | 166 Nm | 130 Nm | 54 Nm | 22 Nm |
| 0.20 s need / vendor peak | 0.98x | 2.7x | 4.5x | 2.4x |
Extending acceleration 0.15 to 0.20 s reduces this optimized torque requirement by roughly 12% torso, 14% shoulder, 18% elbow and 19% wrist, but leaves shoulder/elbow/wrist far above their respective peak ratings. With all four peak torques enforced, the optimistic contact-speed bound is 18.3 m/s at 0.15 s, 20.1 m/s at 0.20 s, 23.9 m/s at 0.30 s, and 31.2 m/s at 0.50 s; by 0.50 s all axes hit the assumed sweep caps and a longer window does not help this constant-acceleration motion. The short-window torque-capped optimizer assigns near-zero torso speed because the distal joints are the bottleneck, not because torso rotation is useless in every trajectory.
Method and limits: a constant-acceleration, same-direction planar 4R sweep from rest; a linear program minimizes the maximum fraction of catalog peak torque to hit 60 m/s and separately maximizes tip speed within peak torque, catalog rated-speed and assumed sweep caps. The mass matrix uses the racket as a 0.323 kg point at its stated 0.330 m balance location from the butt, 0.380 kg wrist motor at the wrist, 0.485 kg elbow motor at the elbow, and 0.960 kg shoulder motor at the shoulder. These point placements are favorable test assumptions, not a measured mechanism. Arm links, torso structure, mounting hardware, motor rotor/gear inertia, racket's distributed moment, impact load, cable losses and compliance are omitted. These outputs are conditional model results, not demonstrated racket speeds or global upper bounds; peak catalog torque is not proven available at the listed rated-speed points. A 3D timed kinetic-chain motion, joint limits, off-axis load paths, stability, thermal/current limits and measured impact are still needed.
Remote-actuation route: putting larger motors near the trunk and transmitting torque to distal joints can remove motor-body mass from the shoulder-to-wrist moving chain. The backdrivable cable-driven Barrett WAM and ESTHER's 10 m/s experiment are real precedent, not proof of ATH's six-times-higher target. A cable design must publish or measure its routing, drum radii, reduction, stretch, friction, tension, fatigue and impact loads before running a credible full-dynamics solve. No catalog-verifiable combination checked here closes the 60 m/s gate.
Next workstream: move the motors off the arm
Remote actuation is the design route, not a verified 60 m/s solution. Mount motor bodies at the base or trunk, transmit joint torque with pre-tensioned cables, and make distal links light. This removes motor-body mass from the moving arm but introduces cable loads, friction, stretch and routing. The first sourced motor-and-ratio example below still fails the target, even under favorable assumptions.
Barrett's WAM inertial specification gives concrete cable hardware: Sava SN2047 1.1176 mm cable around a 9 mm motor drum to a 33 mm intermediate pulley (first stage 3.667:1); a second stage uses SN2054 1.3716 mm cable, an 11 mm intermediate pinion and a 126 mm base pulley, 84.75 mm differential-input pulley or 52 mm elbow pulley. The actual WAM ratios are 42:1 for base, 28.25:1 for shoulder differential inputs and 18:1 elbow, with output-reflected drive inertias 0.205/0.094/0.094/0.0346 kg m² for its first four drives. These are published values for that arm, not a free massless transmission for ATH. The WAM cable-maintenance guide documents pre-tensioning and cable routing constraints. ESTHER measured 10 m/s with a WAM on a powered wheelchair; it does not validate 60 m/s.
Maxon EC-i 52, 48 V, 420 W, part 633919 is a real off-chain motor lead: 752 g body, 1.02 Nm nominal torque at 3990 rpm, 4.83 Nm stall at zero speed, 5090 rpm no load and 170 g cm² rotor inertia. A hypothetical lossless 33/9 = 3.667:1 cable on each axis yields 3.74 Nm nominal joint torque, 17.71 Nm joint stall and 0.000229 kg m² reflected rotor inertia per joint. This is not an assembled or qualified Maxon-WAM drivetrain; a matched encoder, drive electronics, pulley mount, cable rating and safe tension have not been selected. The stall point is not sustainable and cannot occur at high speed.
| Optimistic remote-drive screen | 0.15 s | 0.20 s | 0.30 s | 0.50 s |
|---|---|---|---|---|
| Peak-capped tip speed (m/s) | 11.9 | 15.4 | 20.0 | 27.1 |
Re-use the preceding hypothetical 0.25/0.55/0.30/0.685 m four-axis geometry, assumed 1.5/1.5/1.5/2.0 rad joint sweeps, and 0.323 kg racket mass concentrated at the Blade's measured 0.330 m balance point from its butt. Moving the 0.960/0.485/0.380 kg shoulder/elbow/wrist motor bodies to the base removes 1.825 kg of moving motor bodies; the model's torso/shoulder/elbow/wrist mass-matrix diagonal falls from 1.491/.871/.162/.035 to .661/.450/.128/.035 kg m² before the 0.000229 kg m² reflected rotor term on each joint. (The removed sum is 1.825 kg, not the weight of a finished arm.) A linear torque-speed upper bound joins Maxon's 4.83 Nm stall at zero speed to zero torque at 5090 rpm and includes rotor inertia. It assumes lossless cables, zero link, mount and transmission mass, zero friction, no electrical/thermal limit, no impact and constant same-direction acceleration from rest. The computed speeds are conditional optimistic model results, not actual shots or universal physical limits. Realistic missing loads would reduce them.
Open 60 m/s gate: choose a higher-power motor and compatible controller with a measured torque-speed/current/thermal curve at the actual bus voltage; specify cable or belt diameter, material, pulley radii, strength, wrap, pre-tension, path friction, efficiency, compliance and fatigue; measure link/racket/fixture inertias and base stability. Only then run a time-resolved 3D IK + inverse-dynamics solve through a feasible swing window, check every joint's simultaneous torque/speed/power and cable tension, and validate contact with a real ball/stringbed. No candidate checked here has passed those gates. The same court and ball machine remain the planned sim environment once this design closes on paper.
High-power remote drive: a candidate, not a passed design
No hardware go, even with the revised target. Daniel has said 59.3 m/s is good enough, so the current requested racket-tip threshold is about 59 m/s rather than the earlier 60 m/s. A matched catalog motor and 6:1 gearbox now make a configuration-dependent research solve possible. One restricted-family motion reaches 59.28 m/s at the far racket tip in an idealized planar model, numerically above that revised tip threshold, but not at a selected stringbed contact point or in a qualified drivetrain. The remote joint transmission and real-court hit are not solved.
Teknic ClearPath N143
CPM-SCHP-N1433P-ELSB: at 230 VAC three-phase, the published peak curve is about 32.85 Nm through 1,630 rpm; the continuous curve starts at about 8.10 Nm. Rotor inertia 19.0 kg-cm²; motor body 11.4 kg. Availability of that peak over repeated tennis swings is not established.
Teknic GSLL-120-006-N143B-XX
The N143-input 6:1 gearbox lists 97% efficiency, 230 Nm maximum acceleration torque, 128 Nm nominal output torque, 3.30 kg-cm² input inertia, 6.30 kg mass, 3,000/6,000 rpm nominal/max input and 1,850 N maximum radial load. The motor's 32.85 Nm peak times six and 97% is about 191 Nm: below the gearbox's acceleration rating only, not a validated assembly or duty cycle.
A physics screen, not a tennis hit
The planar four-hinge model recomputes its configuration-dependent mass matrix, velocity terms and racket-point Jacobian during the swing, applies the Teknic peak torque-speed polyline with motor and gearbox reflected inertia, and cross-checks a found motion with a switch-separated RK45 integration (0.1 ms maximum step). It assumes a fixed base, four independent motors, zero gravity, 0.15 s from rest and joint-travel caps of 1.5/1.5/1.5/2 radians. A search varied starting posture, motor direction, torque fraction and one switch per joint. These are found motions in a restricted control family, not maximum speeds or measured shots.
Example moving masses: a hypothetical 0.25/0.55/0.30 m three-link frame made from a published 30/27 mm carbon tube weighing 193.8 g/m, and a Wilson Blade 98 16x19 v9 with listed strung mass 323 g, balance 33.02 cm from the butt, length 68.5 cm and swingweight 324. Using the standard swingweight axis 10 cm from the butt, the parallel-axis calculation gives 0.01528 kg m² about the racket's center of mass. This frame and these link lengths are scenario inputs, not built ATH hardware.
| One found 6:1 swing at 0.15 s | Computed speed | What that point means |
|---|---|---|
| Far tip, 68.5 cm from racket butt | 59.28 m/s | End of the frame, not a selected ball contact |
| 60 cm from butt | 53.17 m/s | Illustrative point, not a chosen stringbed hit |
| 50 cm from butt | 45.97 m/s | Illustrative point, not a chosen stringbed hit |
Architecture variants to test, not selected hardware
A parallel energy store at a proximal joint (for example a torsion or leaf spring) could charge during a longer backswing and release during the roughly 0.15 s stroke. In the one planar, no-ball, fixed-base example above, a re-integration of the found control path has about 410 J peak modeled kinetic energy (about 229 J moving links and racket, 181 J reflected motor/gear rotors), about 555 J gross positive joint work and about 8.1 kW peak summed positive joint mechanical power. These are model-only, path-specific figures, not a measured spring rating or an upper bound. At least 555 J delivered over a 0.5-1 s charge interval would average 1.11-0.56 kW mechanical before charging/discharging losses; recovered energy and the actual spring share are unresolved. A parallel spring could bypass some high stroke torque in a remote cable, but needs a verified spring curve, travel, latch/clutch, controlled release, braking and variable pretension or direction-specific modules so it can aim and vary stroke speed across the whole 3D shot envelope. A 4-8:1 charging tackle or windlass trades cable force for stroke distance and drum speed, not for free energy; cable, pulley and latch loads need ratings. A safe first candidate experiment is one instrumented shoulder-joint spring and charge/release bench rig, measuring torque-angle, stored/recovered energy, peak cable/latch forces, release-time control and repeat-cycle heat before building a whole arm.
A hybrid transmission is another candidate: carry heavy motors at or near the proximal axis, test a short high-force pushrod/four-bar or a rated belt for the elbow, and reserve lower-load cable/belt routing for a wrist only if its measured torque, speed and inertia allow it. For a 200 Nm joint, the tension difference at a 0.10 m effective output radius is 2 kN; a 1:1 remote flexible stage does not remove that joint-side force. A rod can move the high load into compression/tension, but does not make force disappear: its moment arm changes through the stroke, buckling, joint pins/bearings, routing, added mass, stiffness, collision clearance and impact loads must be sized in the actual geometry. The 6:1 motor/gearbox example and unbuilt 1:1 remote path above are not proof of this hybrid. A 15-25 Nm wrist is also just a proposed range, not our wrist requirement; measure it in full shot dynamics. Compare these variants against a sourced base-mounted direct drive, not by rules of thumb alone.
Transmission and court gates remain open
A direct 6:1 cable stage with 0.188-inch Sava 7x19 cable faces a size problem: Sava's 25x-rope-diameter pulley guidance gives at least a 4.7-inch root input pulley and a 28.2-inch output pulley at 6:1. Placing the reduction at the base gearbox avoids that distal size, but a rated near-1:1 cable or belt drive to the joints still has not been selected. At the gearbox's roughly 191 Nm mapped peak, a hypothetical 3.125-inch or 4.7-inch drive pulley needs about 4.8 or 3.2 kN tension difference. Actual shaft radial load needs the cable's directions, wrap, pretension and supports; a scalar tension comparison to the 1.85 kN catalog radial rating cannot prove failure or safety. Cable breaking strength is not a working-load or fatigue rating.
Still missing: real remote-drive rating and packaging, distal pulley and joint inertias, torque losses, compliance, bearing loads and fatigue; motor peak duration and repeated-shot thermal limits; deceleration and regenerated energy; three-dimensional reach, collision-free ball-machine contact and mobile-base stability; a chosen racket stringbed impact location, ball and measured rebound. The existing court and ball machine remain the planned environment; this motion has not been coupled to them. The driven base must cover the court quickly enough for varied pace, spin, bounce and shot positions, not merely hold still for this one arm stroke. Its speed/acceleration target must be set against measured pro court movement and the published ESTHER wheelchair baseline, then checked with braking, turning, grip and arm reaction loads; this arm-only screen sets no base-speed claim. Keep the distal arm light to reduce moving inertia without trading away durability. The assembled carbon-tube links, joints, fixtures and racket mount must be structurally rated for repeated ball impacts at the intended about-59 m/s racket-tip speed, repeated-shot duty cycles and incidental knocks; tube mass per metre alone establishes no strength or fatigue margin. A remote cable drive can provide engineered compliance between impact and gearbox, reducing direct shock into gear teeth, but that benefit is not automatic: measure its stiffness, damping, peak reflected load, tension, backlash/position error and cable fatigue under realistic impacts before selecting it. Optimize structurally rated lever length and light distal mass together: longer reach can add racket-tip speed at a given joint rate, but also adds inertia, bending load, swept volume and clearance risk, making the arm slower to redirect or unable to get out of its own way. Choose length from collision-free, time-resolved three-dimensional shot and recovery solves across the court, not a length heuristic or an unsupported extreme joint-rate assumption. Keep enough controllability to shape chosen ball placement and spin across the shot envelope, not merely reach peak speed; test off-center racket impacts, follow-through braking and regenerative-bus limits, wheel traction and tip stability, carbon-tube end-insert fatigue/delamination, cable bend and snap loads, rally thermal duty and flexing sensor/power harnesses: verify impact-pose and racket-face tracking, motor/transmission control bandwidth, backdrivability or measured reflected impedance, and stability under incoming ball forces. The longer lever also increases bending and joint reaction loads, which must be checked, not assumed away. The revised about-59 m/s racket-tip target must hold across representative three-dimensional swing directions and trajectories, not just the one found planar path: forehands, serves and other shots with varied height and spin. A collision-free, time-resolved IK and full-dynamics solve must reach each selected ball/stringbed contact in the court workspace while the base moves, and verify simultaneous motor torque-speed, transmission loads, stability and shot contact. Until that envelope is checked with real hardware inputs, a single fast trajectory is not evidence of all-direction capability. Despite the revised about-59 m/s racket-tip target, no performance verdict follows from the current screen.
Same ball machine, new player
Build the arm/base simulation by retaining the legacy-serve environment: its court, camera, ball-machine geometry hook __MACHG and MuJoCo feed/ball-physics step __PHYS. Swap only the H2 body and SONIC reference tracker for a wheeled-base plus arm model and a physics-step IK/controller. Publish no fabricated wheel inertia, tire friction, motor torque-speed curve, cooling, racket fixture or stringbed constants. Until those are grounded, show their values as unknown, keep the 60 m/s performance gate failed/unverified, and do not label visual ball returns physically validated. Use the same incoming ball-machine shots to compare the replacement player to the legacy scene.
Parameter ledger
"Published" means a cited component or study number, not a certified value for ATH. Unverified cells cannot silently enter a physics simulation.
| Subsystem | Grounded reference | What remains unverified for ATH |
|---|---|---|
| Arm kinematics | Barrett describes a 7-DoF WAM with approximately 1 m reach. Its published inertial document includes link and drive data but calls itself a draft. Datasheet · inertial draft | Exact selected arm and revision; measured link inertias, joint friction/backlash, end-tool/racket geometry, self-collision and joint-specific torque-speed limits. |
| Motor and drive | HEBI R-series modules and ODrive motor characteristics publish candidate performance data. Quasi-direct-drive (QDD) is a candidate architecture class: a low-reduction transmission paired with a high-torque-density motor, not a selected ATH joint. The MIT Cheetah actuator study explains the tradeoff between reduction, reflected inertia and impact backdrivability. Neither a QDD label nor a component catalog establishes 60 m/s. | Exact motor winding, controller, gearing, voltage, current limits, torque-speed curve under load, cooling and duty cycle. For a QDD candidate, fix motor/winding, drive, voltage, gear ratio and mechanical design, then check published and measured joint torque-speed curves, reflected rotor inertia, electrical power and thermal duration against the full 60 m/s racket-tip trajectory. Joint speed alone cannot establish racket speed. |
| Wheels and court | ESTHER reports driven-wheelchair field measurements and its own drivetrain. Study | Our chassis mass/CG, wheel radius and drive, tire-court grip on each surface, slip, acceleration, braking, tipping and turning with arm motion. |
| Racket and ball | Published stringbed studies measure impact dwell and rebound under stated test conditions, not our racket. Kawazoe et al. · Washida et al. | Chosen ball/racket/string/tension: measured impact map across center, edge and rim; spin transfer, deformation, recoil, wear and high-speed aero/spin decay. Existing site racket contact is not validated. |
| Racket speed target | ATH design requirement: up to 60 m/s racket-head speed. ESTHER measured 10 m/s in its own system, not ATH. ESTHER study | Show a feasible time-resolved 60 m/s tip trajectory through the complete IK chain and measured joint torque-speed/current, reduction, reflected rotor inertia, power, thermal and structural limits. Then measure actual racket-head speed. No selected ATH hardware has passed this gate. |
| Shot and ball-exit targets | Measured impact data must tie a particular inbound ball and racket state to outbound speed, spin and trajectory; tip speed alone cannot set ball exit speed. Impact study · String study | Measure player-specific contact height, racket-head pose and velocity, outgoing ball speed/spin and landing distribution for a Nadal forehand, an Opelka serve and the wider shot set. Test the selected racket/ball stringbed at relevant inbound speeds, impact locations and angles before predicting ball-exit speed at 60 m/s. No "any shot" verdict yet. |
How the model earns trust
- Choose exact hardware revisions, acquire manufacturer curves and measure the assembled arm, base, wheels, racket and ball. Trace each parameter to a source or an experiment.
- Implement full rigid-body dynamics, wheel-court contact, three-dimensional ball flight and spin, stringbed/frame contact and inverse-kinematics shot planning within measured actuator envelopes.
- Check numerical convergence, collisions, energy and impulse accounting. Validate the full racket-head speed time history and measured ball exit speed/spin, rather than deriving either from one peak tip speed. Hold out real trajectories, impact locations, surfaces and shot types before claiming transfer to hardware.
An IK solution is only a reachable pose. It does not prove the arm can get there in time, survive the load or make the ball land. Speed, acceleration, torque, power, tire grip and contact tests must all pass.
Ball machine
Open the site's existing ball-feed simulator to explore launches. It still uses historical H2 assets; it is not an arm-on-wheels simulation or evidence that the proposed robot returns the balls. Its racket contact has not passed calibration.