Product mode: not part of the pre-registered experiment. Simulated in MuJoCo and untested on hardware. A recommendation is a simulation ranking under the stated conditions, not a hardware guarantee.
Build certificate: Floor tiling
Lift a 300x300 mm porcelain tile and set it on the fresh adhesive bed, then press it down until it is within 1 mm of its finished height and no more than 0.5 degrees out of level, using no more than 60 N of pressing force.
multi head Two-tool head: Vacuum area frame + Tamper pad · the Studio's recommendation
Result
| Full-task completion | Runs | Weakest step | Median cycle (with changes) | Fidelity |
|---|---|---|---|---|
| 100% | 8 | lay_tile (100%) | 19.2 s | approximation |
Why: Ties on completion (100%) and the weakest step, with the shorter cycle (19.2 s against 25.2 s).
Setup
| Tool | Effector | Family | Fidelity | Steps |
|---|---|---|---|---|
| 0 | Vacuum area frame, 2 x 2 cups, 300 mm | suction | approximation | lay_tile |
| 1 | Tamper pad, 200 mm | process tool | approximation | tamp_tile |
| Tools | Actuators | Carried (heaviest segment) | Built in total | Changes per run |
|---|---|---|---|---|
| 2 | 5 | 1.408 kg | 1.408 kg | 1 |
| Head | Angle | Standoff | Lateral | Spin |
|---|---|---|---|---|
| 0 | -60 deg | 50 mm | 0 mm | 0 deg |
| 1 | 60 deg | 50 mm | 0 mm | 0 deg |
bracket.stl: two-tool bracket: hub, two arms and two adapters; heads at -60, 60 deg, 50 mm standoff; 135 g printed; watertight single solid: True.
Per step
| Step | Primitive | Object | Tool | Success | Failures | Fidelity | Caveats |
|---|---|---|---|---|---|---|---|
| lay_tile | grasp (as suction) | floor_tile | 0: Vacuum area frame, 2 x 2 cups, 300 mm | 100% | none | approximation | The mortar bed is an approximation: nine plastic cells that yield and harden linearly and keep their set; no flow, suction or curing. A screen, not a promise. Suction cup (approximation): a compliant lip of rigid segments on a bellows spring with a limited tilt; the seal is a geometric test (every lip segment touching the face, part material under 13 rays inside the lip) gating a MuJoCo adhesion actuator capped at vacuum x lip area (at most 60 kPa). The lip must touch the face: no pull across a gap. Not modelled: air flow, leak rate, porosity, vacuum build-up time (the vacuum follows the shared close ramp and is cut the control tick the seal is lost), lip material hysteresis, peel dynamics beyond the contact distribution. The lip's friction is the part's. Treat suction results as a screen, then verify on hardware. |
| tamp_tile | push | floor_tile | 1: Tamper pad, 200 mm | 100% | none | approximation | The mortar bed is an approximation: nine plastic cells that yield and harden linearly and keep their set; no flow, suction or curing. A screen, not a promise. Process tool (approximation): a rigid tool used by the Studio's own twist or push primitive. The nut runner's spindle is the wrist's roll through a torque-limiting clutch [assumed]; the tamper's stage is a spring that slips at its overload force. The torque, the force and the contact are simulated; the process (thread cross-threading, bed compaction) is not modelled beyond them. Verify on hardware. |
Conditions
Every compared setup ran the same conditions and seeds in the same order. A seed draws the friction jitter (+-10%) and a +-0.5 mm aim jitter; the placement errors below are hidden from the controllers.
| Condition | Seed | Placement error x/y (mm) | Run | Cycle | Trajectory |
|---|---|---|---|---|---|
| nominal | 0 | none | completed | 19.2 s | 1a71f3c7503720e6 |
| place01 | 1 | floor_tile +0.31/-0.19 | completed | 19.2 s | 293a6f5f1c13a05d |
| place02 | 2 | floor_tile +0.09/+0.24 | completed | 19.2 s | 195802099f0bb138 |
| place03 | 3 | floor_tile -1.50/-1.83 | completed | 19.2 s | 89d686b25acd366b |
| place04 | 4 | floor_tile -0.00/+1.47 | completed | 19.2 s | 47d3c33dd7dadc8d |
| place05 | 5 | floor_tile -1.63/-1.20 | completed | 19.2 s | f912b9c3daf47759 |
| place06 | 6 | floor_tile -0.23/+0.27 | completed | 19.2 s | 2842860a4df12b45 |
| place07 | 7 | floor_tile -1.50/+1.48 | completed | 19.2 s | feca2c3a237bc9f6 |
Compared setups
| Setup | Kind | Completion | Weakest step | Cycle | Tools | Eligible |
|---|---|---|---|---|---|---|
| Two-tool head: Vacuum area frame + Tamper pad | multi head | 100% | lay_tile 100% | 19.2 s | 2 | yes |
| Tool change: Suction cup > Tamper pad | tool change | 100% | lay_tile 100% | 25.2 s | 2 | yes |
| Two-tool head: Suction cup + Tamper pad | multi head | 0% | lay_tile 0% | 9.2 s | 2 | never completed the task (stops at lay_tile in 8/8 runs: criterion_not_met 8x) |
| Two-tool head: Suction cup + Tamper pad | multi head | 0% | lay_tile 0% | 9.2 s | 2 | never completed the task (stops at lay_tile in 8/8 runs: criterion_not_met 8x) |
Sensitivity to change times
Re-ranked with tool changes of 3, 6 and 12 s and head switches of 0.75, 1.5 and 3 s: the recommendation is the same for every pair; the order changes in 0 of 15 pairs.
Assumed constants
Equal for every candidate; none is sourced from a target arm or changer.
| Constant | Value | Note |
|---|---|---|
| roll_torque_nm | 15 N m | wrist roll torque limit |
| roll_speed_max_deg_s | 180 deg/s | wrist roll speed cap |
| pitch_range_deg | 90 deg | wrist pitch limit, +- from straight down |
| pitch_speed_dps | 180 deg/s | wrist pitch speed cap |
| pitch_torque_nm | 15 N m | wrist pitch torque limit |
| tool_change_s | 7.5 s | automatic tool change time (a fixed clock advance) |
| present_s | 1.5 s | multi-tool head switch time (a fixed tilt, not simulated) |
| changer_master_kg | 0.25 kg | tool changer master plate (stays on the wrist) |
| changer_tool_plate_kg | 0.1 kg | tool changer plate on every tool |
| changer_stack_mm | 52 mm | tool changer stack height |
| arm_payload_kg | 3 kg | largest tool stack the arm may carry; heavier setups are refused |
| bracket_density_kg_m3 | 1300 kg/m^3 | printed carbon-filled polymer bracket |
| placement_error_mm | 2 mm | per-object placement error of the condition grid (uniform) |
| simplicity_per_extra_effector | 0.05 rate | integrator rule R01: completion a setup must add per added effector or changer |
| simplicity_step_floor | 0.5 rate | integrator rule R01: a weakest step under this fails outright |
| tie_margin | 0.05 rate | completion-rate tie among setups of the same complexity |
| cycle_margin | 0.1 fraction | among setups of the same complexity, a simpler one may be this much slower |
| soft_arm_max_part_kg | 2.5 kg | integrator rule R09: the heaviest part a soft arm may carry |
| hook_min_mm | 5 mm | a rigid tool hooks if its tip reaches this far sideways past the tool and flange |
| hook_throat_mm | 6 mm | room a lip needs between a hook and the flange, else the flange blocks it |
| process_torque_nm | 4 N m | integrator rule R05: tightening a threaded member to this torque or more needs a nut runner (1 N m above hand-tight) |
| socket_fit_mm | 0.5 mm | a nut runner's socket drives a hex at most this much under its own size |
| sweep_clearance_mm | 5 mm | multi-tool head swept-volume clearance |
Caveats
- The mortar bed is an approximation: nine plastic cells that yield and harden linearly and keep their set; no flow, suction or curing. A screen, not a promise.
- Suction cup (approximation): a compliant lip of rigid segments on a bellows spring with a limited tilt; the seal is a geometric test (every lip segment touching the face, part material under 13 rays inside the lip) gating a MuJoCo adhesion actuator capped at vacuum x lip area (at most 60 kPa). The lip must touch the face: no pull across a gap. Not modelled: air flow, leak rate, porosity, vacuum build-up time (the vacuum follows the shared close ramp and is cut the control tick the seal is lost), lip material hysteresis, peel dynamics beyond the contact distribution. The lip's friction is the part's. Treat suction results as a screen, then verify on hardware.
- Process tool (approximation): a rigid tool used by the Studio's own twist or push primitive. The nut runner's spindle is the wrist's roll through a torque-limiting clutch [assumed]; the tamper's stage is a spring that slips at its overload force. The torque, the force and the contact are simulated; the process (thread cross-threading, bed compaction) is not modelled beyond them. Verify on hardware.
- The floating wrist abstracts the arm: there is no arm reach or arm collision model.
- Presenting a multi-tool head is a fixed tilt per segment plus a fixed time, not a simulated motion.
- A tool change is a fixed time, not a simulated docking.
- Controllers are scripted and open-loop on the placement error.
- A recommendation is a simulation ranking under the stated conditions, not a hardware guarantee.
Provenance
| Item | Value |
|---|---|
| Contract | rtt-effectors/2 |
| Task id | f9553b22ce12fc141c274d5532f25c97083ff5cd41faa49fc1927d9487fbc5a3 |
| Setup id | 4adcc3231c74df612f82b2dbaba5f6dda3ab6ce38843e5a56d3752960e8519b0 |
| Fairness hash | 82b5bcdac3b461e77bc17444036a535a9acb32f29bafd804947228e68090636d |
| Controllers | default |
| Screen | 132 runs on 3 conditions |
Product mode: not part of the pre-registered experiment. Simulated in MuJoCo and untested on hardware. A recommendation is a simulation ranking under the stated conditions, not a hardware guarantee.