DESIGN. TEST. UNDERSTAND.
Explore how a mobile chassis, a 1.6 m arm and four moving cameras work together to intercept a shuttle. Every attempt is here—successes and misses.
A project by Agasthya & Rajesh
A computational engineering study, shared for discussion and teacher feedback.
Read the experiment’s boundaries ↓What passed: a finite, independent single-return simulation test. Still to test: return-to-home recovery and continuous rallies. Build and purchases remain on hold.
01 / WATCH THE ACTUAL TEST PATHS
Each attempt starts at home. Automatic playback moves between separate tests; it does not demonstrate a continuous rally. After 500 ms of recorded braking, the robot pose is held while the calculated outgoing shuttle finishes its flight.
The recorded datasets use feather shuttle physics. Nylon has not been tested.
02 / SEE THE WHOLE SAMPLE
Each point shows the chassis position at contact or closest approach. Success is shown by a circle; a failed attempt by a cross. Select a point to inspect its recorded replay.
Select a point to open the replay.
Coordinates in metres: net x = 0; back line x = −6.70; sidelines y = ±2.59. Racket paths are projected onto the floor; use the oblique replay to inspect height.
The overall score includes every setting, including the weakest ones.
03 / CONNECT THE DRAWING TO THE TEST
Alternative viewpoints allow another pair to track when a head is blocked. Mounting on the fixed wheel structures removes the tall camera posts that crossed the arm’s path. Camera pivots are 0.50 m high; the pan and tilt motors move each head independently.
During a stroke, front cameras turn into a restricted rearward sector to avoid the arm. Rear cameras do much of the tracking. Four heads do not guarantee four simultaneous views.
New delayed observations arrive while the chassis and arm are already moving. Every 10 ms, the controller updates its estimate and commands against a frozen library of 21 feeder settings. Motor torque, grip and actuator response are represented in MuJoCo.
This is a finite feeder controller, rather than a general AI that has learned to play a human. Practice and broader trajectory prediction are later stages.
THE EXPERIMENT’S BOUNDARIES
The animation is a schematic drawn from recorded coordinates. The design gallery shows the actual native CAD. Display interpolation is for viewing; it does not replace the 1 ms dynamics integration or the saved 5 ms traces.
Research inspiration: published robot study, arXiv 2504.17771. Our 21 synthetic feeder settings are our own test definition; they do not reproduce the authors’ exact machine settings.
04 / A BETTER DESIGN STARTS WITH QUESTIONS
Help Agasthya decide what to test next.
The notes stay in this browser. Export them as a file to share manually; nothing is sent to a teacher or an online service.