Linewise · BigQuery VLM + VL1 · 28 Aug 2026

Stationary gripper-only task coverage

Stationarity is judged over the complete repeatable work cycle, not a cropped manipulation moment. The robot base stays fixed; every routine input, output, reset, control, and handoff is inside one bounded reachable workcell or is locally presented by a fixture, bin, rack, or conveyor. Walking, driving, carrying between stations, and mobile-base repositioning disqualify the normal cycle. The same stock parallel two-finger gripper remains installed throughout.

70gripper-only task families
350concrete gripper operations
11scored high confidence
3mobile or unsupported cycles excluded

Gripper task suitability

All rows use a fixed robot base and stock two-finger jaws. SINGLE one gripper · DUAL two stock grippers · Direct grasp jaws contact the workpiece · Gripper-held item jaws hold an ordinary passive task item. Ranked by confidence; expand a row to see its whole-cycle stationarity class, multiplier, and note.

Showing 70 of 70 task families

Task familyGripper setupGripper confidence Top environment types

Scope and scoring

The normal repeatable cycle is classified before scoring; a short stationary excerpt from a mobile workflow is not enough.

  • Complete-cycle stationary gate. The robot base and workcell stay fixed, and every routine per-item input, output, reset, control, task item, and handoff fits inside one bounded reachable envelope. Inputs and outputs may be locally presented by a fixture, bin, rack, conveyor, or other material system. Human setup or bulk replenishment may happen outside the robot task, but ordinary cycle completion may not require walking, carrying between rooms or stations, moving a cart, or repositioning a mobile base.
  • Stationarity classes. Fixed cycle receives 1.00 when the evidence and scoped operation support a complete local cycle. Local feed required receives 0.85 when the operation is valid only with explicit in-cell presentation and removal. Fragment-only evidence receives 0.45 when footage shows a locally stationary manipulation but does not establish that the surrounding workflow is stationary; this cap keeps the family Low even with a perfect manipulation prior. Mobile cycle receives 0.00 and is omitted from the scored catalog.
  • End-effector gate. Every active arm must end in the same stock parallel two-finger gripper for the whole task. Powered tools, mounted utensils, custom tool heads, suction cups, dexterous hands, and replacement end-effectors are excluded.
  • Direct grasp versus gripper-held item. In a direct task, the jaws contact the workpiece. In a gripper-held-item task, the jaws grasp an ordinary passive item by the same handle or material a person would hold; the item is never mounted to the wrist and the gripper is never removed. These rows receive a grasp-stability penalty.
  • Single versus dual. Single needs one stock gripper. Dual needs two arms, each ending in the same stock gripper, for holding-open, tensioning, or opposing-pull actions.
  • Manipulation prior (judgment). Whether the specified stock-gripper setup can physically perform the task. Rigid objects, visible grasp faces, short transfers, and fixture-guided placement start highest. Deformable materials, fragile objects, regrasping, sustained contact force, and sub-millimetre alignment are penalized. It is a major score input after the stationarity gate and is an estimate, not a measurement.
  • BigQuery formula. 100 × manipulation prior × stationarity multiplier × [0.78 + 0.22 × mean confidence × (1 − 0.5 × ambiguous rate)]. Weak labelling drags a score down; it cannot rescue a mechanically implausible or mobile task.
  • VL1 formula. VL1 records do not expose calibrated confidence or ambiguity fields, so their evidence term is the documented 0.92 factor: 100 × manipulation prior × stationarity multiplier × 0.92.
  • Embodiment caveat. The footage labels show that the task and environment exist in the catalog; they do not prove that the recorded human workflow fits a fixed robot envelope or that a robot gripper has executed it successfully. Each deployment still needs workcell-layout and hardware validation.
  • Bands. High ≥ 72, Medium 50–71, Low < 50. High is the strongest near-term fit. Medium is feasible only with the stated setup and constraints. Low should not be offered without a hardware trial.

Hard-excluded source families after the whole-cycle audit:

  • Wheel & tire handling at a fixed service bay. The normal service cycle spans wheel storage wheel-service equipment and multiple vehicle hubs or requires vehicle or base repositioning.
  • Fixed-fixture pruning with manual shears. Available evidence is landscape hedge and field pruning whose normal cycle advances across plants; it does not show repeated fixture-presented pruning.
  • Fixed-workcell handled-bag staging. The two matched assets show distributing prepacked bags across event tables rather than a repeated fixed-cell transfer cycle.

The catalog is intentionally split into concrete capability slices. Broad footage families such as pick-and-place or small-parts handling support several sellable gripper operations: packing, transfer, kitting, fixture loading, hardware placement, and guided insertion. Slices can overlap and can inherit evidence from the same broader source family, so asset counts and hours must never be summed across rows.

The remaining gap is geometry and reach: the labels do not measure object size, mass, friction, required jaw clearance, grip force, wrist torque, or the distance between every source and target pose. A reach-envelope check, grasp-feasibility pass, and hardware trial are required before turning a score into a deployment commitment.