Power Tool Use Demonstrations

Why power tools are a separate sourcing problem

A hand tool such as a wrench transmits only what the person applies. A drill or impact driver adds a motor, a clutch or hammering mechanism, and a trigger, so the tool pushes back. The operator braces against reaction torque, feels the clutch ratchet, and hears the motor load change when a screw seats. A model that watches only the visual stream misses most of that.

Safety also changes what can be recorded. Operators wear eye protection and often gloves, work at awkward angles, and hold a workpiece with the other hand. Those facts affect camera placement, occlusion and who is willing to be filmed. The sibling page on hand tool use demonstrations covers the unpowered case; this page covers what powered operation adds.

If you are still deciding whether footage is enough, the maintenance cluster hub and the humanoid data sourcing guide place this task among its neighbors.

Task anatomy

A single "drive a screw" instance contains more steps than the clip title suggests. The steps below are the ones annotators usually need to mark.

1. Select the tool and bit; check the bit is seated and the direction switch is set.
2. Set clutch or speed range, where the tool has them.
3. Position the fastener on the workpiece and align the bit with the head.
4. Start at low trigger pressure, then modulate to drive.
5. Seat the fastener: the clutch clicks, the motor slows or an impact mechanism engages.
6. Release the trigger, withdraw straight, and check flushness or torque by hand.
7. Recover from cam-out, a stripped head, a cross-threaded start or a jammed bit.

Drilling replaces steps 4 to 6 with entry, feed pressure, chip clearing and breakthrough. Breakthrough is the moment of highest kickback risk and is worth its own label.

Objects vary across fastener type (wood screw, machine screw, anchor), head drive (Phillips, Torx, hex), material (softwood, sheet metal, drywall) and tool format (corded, cordless, right-angle, impact driver). Coverage across these axes matters more than clip count, because each combination changes the contact dynamics.

What published work suggests about structuring tool-use data

Tool-use data is easier to learn from when the tool, the action and the target are labeled separately. TACO contains 2.5K bimanual motion sequences across tool-action-object combinations, with 12 third-person views plus one egocentric view, hand-object 3D meshes, 2D segmentation and behavior triplet annotations (Liu et al., 2024). It uses optical motion capture inside an acquisition pipeline, which is a purpose-built setup rather than archive footage.

For the physical side, RH20T reports more than 110,000 contact-rich real-world manipulation sequences with visual, force, audio and action data per sequence, each robot sequence paired with a human demonstration video and a language description (Fang et al., 2023). The pairing is the useful pattern for buyers: human video for understanding, robot-side signals for control. A contractor's clip supplies only the first half.

What the data can support, and what it cannot

Human power tool video does not contain robot actions. Turning it into control data requires retargeting, inferred actions or co-training with robot data, which the embodiment gap page explains. A humanoid also needs a hand and wrist that can hold a tool against reaction torque, which is a hardware question the dataset does not answer.

Limitations to plan for

• Audio is useful but fragile. Motor and clutch sound can indicate seating, but shop noise, hearing protection microphones and wind noise degrade it. Record a clean reference channel if you intend to use sound.
• Gloves and tool bodies occlude hands. Grip labels from video alone are unreliable on closed-fist grips.
• Trigger state is invisible. Without an instrumented tool or a tracked finger, trigger depth must be inferred from motion.
• Torque settings are rarely logged. Existing clips seldom record clutch setting, battery level or fastener specification.
• Happy-path bias. Experienced workers rarely stage stripped heads or cam-out. Recovery steps usually require scripted failures.
• Safety content may be sensitive. Footage that shows bypassed guards or missing protective equipment can create obligations for the supplier.

Existing footage or a scripted capture

Existing material comes from trade-training videos, contractor body cameras, workshop process recordings and apprenticeship archives. It is fast to review and shows real conditions. It rarely has consistent framing, audio quality or labels, and consent was usually given for training or safety purposes. The trade-offs are discussed in existing workplace footage versus commissioned recording.

A scripted capture designs those properties in: fixed fastener and material grids, a chosen camera layout, synchronized audio, a calibration routine and explicit consent for AI training. It costs more and needs site access, workers willing to be recorded, and a safety plan for the recording itself. For force and trigger signals, see instrumented tools and tool pose tracking.

Illustrative capture specification

# Illustrative only - synthetic values, not an available dataset
task_family: fastener_driving_and_drilling
tools: [cordless_drill_driver_A, impact_driver_B]   # model, battery, clutch setting logged
fasteners: [wood_screw, machine_screw, self_tapping_screw]
materials: [softwood_board, sheet_metal_panel]
cameras:
  egocentric: 1 head-mounted, 1080p, constant frame rate
  third_person: 3 fixed, time-synchronized, calibrated
audio: shotgun mic + wearer mic, shared clock
labels:
  steps: [select, set, position, start, seat, release, recover]
  tool_bit_fastener: per clip
  trigger_state: on/off (annotated) or depth (if instrumented)
  outcome: seated | stripped | cam_out | missed | abandoned
  failure_injection: flagged where scripted
metadata: site_017, operator_A12 (pseudonymous), ppe_worn, glove_type
rights: AI-training consent from each operator; no redistribution

Acceptance criteria you can set

• Fastener head and bit tip visible in at least the share of seating frames you specify.
• Step boundaries agree between two annotators on a sampled subset, with your chosen tolerance in seconds.
• Audio and video offset measured against a clap or LED reference and reported per session.
• Failure cases present at the proportion your scripting plan states.
• Held-out evaluation split defined by operator or site, not by random clip.

Recording constraints specific to powered tools

Camera placement needs more thought than for hand tools. A head-mounted camera sees the bit tip only until the tool body blocks the view, and a fixed side camera loses it when the operator turns away. Plan at least one view looking along the tool axis and one from the side, and check how both behave in confined spots, which the page on tool use in tight spaces treats separately.

Dust, chips and sparks foul lenses, so ask how often lenses are cleaned and whether clips with obscured frames are flagged. Battery tools also change behavior as the pack drains, so log battery state at the start of each session. Finally, agree in advance what the recording crew does when a worker takes an unsafe shortcut: stop the session, continue without comment, or exclude the clip. Each choice shapes what the dataset teaches.

Questions to ask suppliers

1. Which tool models, clutch modes and battery types appear, and are they logged per clip?
2. Is audio recorded, and was hearing protection or ambient noise controlled?
3. Do clips show the full approach, drive and release, or only the visible portion?
4. Are failed starts, stripped heads and jammed bits included or edited out?
5. What camera positions were used, and do they stay fixed across sessions?
6. Did each worker consent to AI training use, or only to training and safety review?
7. Does any footage show bypassed guards or missing protective equipment, and how is that handled?
8. Can the supplier add an instrumented tool or synchronized torque trace if you need one?

Discuss your requirements

If you need drilling or driving data, describe the tools, materials, viewpoints and signals your project requires. SourceX reviews the requirement and follows up about potential sources among US businesses; whether a commissioned recording is feasible is discussed case by case. A request does not guarantee a match. The related catalog entry on first-person video of manual work describes existing-footage routes, and the torque and fastening specification page covers the text side of fastening.

Describe the workplace demonstration data your robotics project needs

Sources

• TACO: Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding — tool-action-object annotation structure and multi-view capture.
• RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — force, audio and action data paired with human demonstration video.