NBtools includes short clips cut from public YouTube videos (videos/raw, videos/frames_10fps) and data derived from DexYCB. The videos belong to their owners and are provided for non-commercial research only; do not re-distribute them. Data licence: CC BY-NC 4.0.

NBtools: where the hand goes on a power drill, a glue bottle, a spoon, a fork, a claw hammer and two hand brushes

Seven tools, each as real meshes with per-vertex hand-contact annotations:

card	what is in it	key frames
`035_power_drill`	11 real cordless drills (DeWalt DCD709 / DCD780 / DCD799, Hilti SFC 22-A / Nuron SF 4H-22 / SF 6H-A22, Bosch GSB 18V-65, Metabo HPT DS18DDX, Milwaukee 2902-20, Craftsman CMCD700, Black+Decker LDX120C), each with its own mesh at the maker's size, synthesis templates per use, a shape family; the YCB drill itself (`mesh/`) has templates for pick up / hand over (5 DexYCB takes, measured), drilling, holding, place and carry (`templates/`)	`keyframes/`: 21 (14 web-video frames, one per video and use, each verified by overlay review + independent critic; 5 DexYCB takes, one key frame each, left-hand takes mirrored to the right; 2 camera-only estimates); `keyframes_v54_archive/`: the 67 earlier frames as they were (its 42 web-video frames were re-run through the current pipeline: 30 pass, and the best one per video and use is in `keyframes/`)
`x03_bambu_liquid_glue`	Bambu Lab liquid glue bottle: grasp, open cap, press sponge, spread, close cap, plus carry / hand over / place	`keyframes/`: 10 filmed key frames + 32 derived shape / symmetry variants
`031_spoon`	YCB spoon: eating soup, scooping, stirring, scraping, skimming, cutting soft food, pick and place	`keyframes/card/`: 646 (104 key frames: 2 web video, 34 TACO and 68 OakInk2 motion capture; 542 sampled variants of them; 35 of the 646 are pre-grasp moments, `phase: pregrasp`)
`030_fork`	YCB fork: piercing food, stirring, holding, pick and place	`keyframes/card/`: 39 (7 key frames: 2 web video and 5 OakInk2 motion capture; 32 sampled variants)
`x10_workpro_hammer`	WORKPRO 8 oz claw hammer: hammering (striking with the head)	`keyframes/card/`: 205 = 192 hold (7 TACO key frames; 116 local variants, 69 wrist shifts) + 13 pre-grasps at pick-up (`phase: pregrasp`, not yet reviewed)
`x11_nacena_brush`	NACENA flat hand broom: sweeping / dusting	`keyframes/card/`: 1,739 = 1,693 hold (33 TACO key frames; local variants, wrist shifts and 180-degree symmetry copies) + 46 pre-grasps (not yet reviewed)
`x12_zihan_brush`	long-handled scrub brush (the pi2brush deployment mesh, 330 mm): brushing / dusting	`keyframes/card/`: 710 = 679 hold (34 TACO key frames; local variants, wrist shifts) + 31 pre-grasps (not yet reviewed)

Units are metres. Every array lives in its mesh's object frame and follows that mesh's vertex order.

Download

The whole dataset (with the videos) is about 1 GB. To get everything:

pip install -U huggingface_hub
hf download henrybao135/NBtools --repo-type dataset --local-dir NBtools

To get one tool, or one drill model (mesh, templates, family and its key frames):

hf download henrybao135/NBtools --repo-type dataset --local-dir NBtools --include "x03_bambu_liquid_glue/*"
hf download henrybao135/NBtools --repo-type dataset --local-dir NBtools --include "031_spoon/*" "030_fork/*" "keyframes_index.json"
hf download henrybao135/NBtools --repo-type dataset --local-dir NBtools \
  --include "035_power_drill/models/dewalt_dcd799/*" "035_power_drill/keyframes/dewalt_dcd799/*"

To get the archived 67-frame drill card (keyframes_v54_archive/), with the meshes its frames refer to:

hf download henrybao135/NBtools --repo-type dataset --local-dir NBtools \
  --include "035_power_drill/keyframes_v54_archive/*" "035_power_drill/mesh/*" "035_power_drill/models/*/mesh/*" "keyframes_index.json"

To keep only the archived frames that passed review, filter keyframes_index.json on set == "keyframes_v54_archive" and review_accept == true, then download just those folders:

import json
from huggingface_hub import hf_hub_download, snapshot_download
ix = json.load(open(hf_hub_download("henrybao135/NBtools", "keyframes_index.json", repo_type="dataset")))
good = [r["path"] + "/*" for r in ix if r["set"] == "keyframes_v54_archive" and r["review_accept"]]
snapshot_download("henrybao135/NBtools", repo_type="dataset", local_dir="NBtools", allow_patterns=good)

The same from Python:

from huggingface_hub import snapshot_download
root = snapshot_download("henrybao135/NBtools", repo_type="dataset",
                         allow_patterns=["035_power_drill/models/dewalt_dcd799/*", "035_power_drill/keyframes/dewalt_dcd799/*",
                                         "keyframes_index.json"])

Where each key frame's affordance is

Each key frame has its own folder: <card>/keyframes/<tool model>/K<nn>_<kind>_<clip>_<frame>/. <tool model> is a drill model, or card for the card's own mesh. All key frames are listed in keyframes_index.json.

file	content
`affordance.json`	card, tool model, use, time, kind, source frame, digits in contact (vertex counts), `status`, `review` (verdicts and notes), `mesh` (the mesh path, relative to the card folder)
`affordance.npz`	`digit`: per tool vertex, 0 palm, 1 thumb, 2 index, 3 middle, 4 ring, 5 little, 255 none (uint8), as reviewed `digit_tips` / `digit_links`: the same map as fingertip pads only / whole fingers (middle and proximal links added along the surface toward the palm) `contact_mm`: tool surface to hand distance, clipped to -10..40 mm `hand_vertices`: 778 x 3 MANO, object frame `hand_contact_mm`: hand surface to tool distance
`fingers.ply`	the tool mesh coloured by digit; open it in any mesh viewer (MeshLab, Blender, trimesh) to see each finger
`contact.ply`	the tool mesh coloured by contact (light to red = touching)
`hand.obj`	the MANO hand mesh of this key frame
`frame.jpg`	the key frame image (video or DexYCB frame)
`../<clip>.mp4`	the raw clip this key frame comes from (one level up, shared by the key frames of that clip)

status says how far to trust a frame:

verified: marked fingertips + expected regions; overlay review + independent critic: the drill's pose was fitted in the frame, every visible fingertip was marked on the frame and cast onto the mesh (measured), a finger hidden on this frame but seen within 1 s of the same grip takes that spot, and a finger never seen gets the region its use expects (e.g. middle / ring / little in thirds of the grip below the trigger). affordance.npz["provenance"] says which: 1 measured, 2 neighbouring frame, 3 expected region; affordance.json["digits_source"] lists the fingers per source. No hand mesh. When the frame cannot show which side of the grip the thumb is on, no left / right side region is drawn.
verified: x-ray fit + overlay review + independent critic: the hand was fitted in that frame's camera, hidden fingers were predicted from the grasp for this drill and use, and two reviewers checked the overlay on the full-resolution frame.
as shown on the page: DexYCB-measured or estimated frames (drill), and the filmed glue frames. The glue frames were released by the owner's review; review keeps the overlay reviewer's notes (fingers on the far side not drawn in several frames).
derived from a filmed key frame: glue shape and symmetry variants (rotations, shifts, turns) of a filmed grasp.
verified: marked fingertips + fitted MANO hand; overlay review + independent critic (spoon / fork web video): fingertips marked on the frame, a MANO hand fitted to them and to the 2-D hand joints with a fitted scene scale (real cutlery and monocular depth differ from the YCB mesh), then a real-size hand re-settled on the YCB mesh at the same contacts; video_utensil_scale records the fitted scale.
measured: TACO / OakInk2 mocap hand at its key frame, transferred to this mesh at the same contacts; ... (spoon / fork): a motion-capture hand at the middle of its longest stable hold, the dataset tool aligned to the YCB mesh (scaled ICP), a real-size MANO hand re-settled at the carried contacts. affordance.json["measured"] names the take and the alignment error.
derived from a reviewed key frame (sampled...); 3-D plausibility judge + critic (spoon / fork): a sampled variant of a key frame (slide along the handle, a small roll, wrist pitch / yaw, finger curl; variant_params; the wrist stays within 15 degrees of the key frame's), re-settled with no penetration and every touching pad within 2 mm. Kept only if a 3-D judge AND an independent critic both call it a plausible hold. Every spoon / fork key frame passed the same 3-D check. Where a key frame failed it, only its passing variants are released (variant_kind ends in seed hand rejected).
archived v54 key frame (not re-verified): the earlier 67-frame drill card. review.accept is the 10-20 overlay verdict when that frame was reviewed.

Fingertips or whole fingers

Every map comes in two views. Fingertips: each finger's distal pad (the part of its area within about 25 mm of its point farthest from the palm, along the surface). Whole fingers: the finger's middle and proximal links added along the surface path from its tip toward the palm contact, about 18 mm wide, up to the finger's length. Web key frames measure fingertips, so their whole-finger area is grown, not measured; DexYCB and fitted MANO hands measure whole fingers, so their fingertip view is the cut-down one. Key frames: affordance.npz["digit_tips"], ["digit_links"]. Templates: p_digit_tips / contact_tips, p_digit_links / contact_links next to p_digit / contact. Card templates (use_templates.npz): <id>__digit_tips, <id>__digit_links.

Spoon and fork

031_spoon and 030_fork use the YCB meshes (mesh/tool.obj). Every grasp is one folder in keyframes/card/. Filter keyframes_index.json on source:

web video: YouTube frames that passed both the video-overlay review and the 3-D check.
TACO: kitchen spoons up to 10 cm wide; ladles were left out.
OakInk2: an eating spoon and fork close to the YCB size.

Filter on variant too: null means the key frame itself, otherwise the sampled variant's name. Phase: phase in affordance.json and keyframes_index.json is hold (the hand holds or uses the tool) or pregrasp (35 spoon grasps from 6 key frames: the moment of reaching, touching or closing on the handle while the spoon still rests on the table; kept for pre-grasp / reach studies, filter them out for stable grasps). 3 TACO spoon key frames with no usable camera view are left out. Variants keep the tool use of their key frame (10-25 release; the 10-24 release had mirrored variants that turned the wrist 40-60 degrees and were removed). A tool symmetry is used only when it maps the working part (bowl, tines) onto itself with its open side unchanged: a rotation then gives a right-hand copy, a mirror a left-hand grasp (not released). The spoon's only symmetry is its mirror plane (left hand only); the fork's 180 degree roll and flat-plane mirror flip the tines and are not used. So every variant is a local move from its key frame: slide along the handle up to 25 mm, roll up to 10 degrees, wrist pitch / yaw up to 8 degrees, finger curl up to 15 %, re-settled with the hand orientation held near the key frame's and rejected when the wrist turns or the forearm swings more than 15 degrees (median 8.9).

Grasp type: affordance.json["grasp_type"] (and <card>/grasp_types.json with the contact features) gives the Feix taxonomy name, primary first and an optional second, from the fitted hand's pad contacts, palm, wrap and index posture. Spoon: Tripod 491, Quadpod 69, Prismatic 4 Finger 48, Light Tool 26, Index Finger Extension + Tripod 6, Small Diameter 4, Tripod + Writing Tripod 2. Fork: Tripod 24, Quadpod 7, Palmar Pinch 7, Tripod / Writing Tripod 1. Rule-based: on stratified samples a judge and a critic agreed with 27 and 31 of 36 primary labels (all fork labels right); the rare spoon labels (Prismatic 4 Finger, Small Diameter) are the least reliable.

Their fingertip view (digit_tips) is the fitted hand's distal pads within 5 mm of the surface. The whole-finger view (digit_links) is every hand part within 12 mm. segment_links labels each whole-finger vertex with the finger segment touching it: 1 tip, 2 middle, 3 proximal, 0 palm, 255 none. Templates are pooled per use in templates/<use>.npz (p_digit, p_digit_tips, p_digit_links, instances). Variants are augmentation of their key frame, not new observations.

Claw hammer and hand brushes

x10_workpro_hammer (WORKPRO 8 oz claw hammer), x11_nacena_brush (NACENA flat hand broom) and x12_zihan_brush (a 330 mm long-handled brush, the mesh of the pi2brush robot deployment) use TACO (Liu et al., CVPR 2024, CC BY 4.0) hammering / brushing / dusting takes. Hold key frame = the middle of the longest stable hold (thumb + 2 fingers within 4 mm); TACO's tool scan is aligned to our mesh (scaled ICP), each contact goes to the nearest surface point facing the same way, and a real-size MANO hand is re-settled there. Every hold seed and variant passed a 3-D plausibility judge AND an independent critic (held from opposite sides, natural, semantics below).

Semantics (affordance.json["semantics"]): hammer on the handle with the head free, and no hand point past the striking-face plane with the handle raised 5 degrees at impact (striking_face_clearance_mm); brushes on the handle or the back / sides of the head, bristles never touched.
Variants (variant_kind): local (slide up to 25 mm, roll up to 10 degrees, wrist pitch / yaw up to 8 degrees, finger curl up to 15 %, wrist within 15 degrees of the key frame), wrist shift (the whole hand moved as one piece along the handle over every position where all contacts stay on allowed parts, and up to 12 mm toward the fingertips or the palm; orientation unchanged) and, for the NACENA broom only, symmetry (a 180-degree turn that keeps the bristles down: the same grasp, still a right hand). The hammer and the long brush have only mirror symmetries (left hand), so no copies.
Pre-grasps (phase: pregrasp, variant_kind: pregrasp, name pre2_f<frame>): the pick-up moment, the last TACO frame before the tool lifts in which the thumb and at least one finger are within 8 mm of the handle (more fingers preferred), transferred and settled like a hold seed and kept only if the thumb faces a finger across the handle. They are filed under their take's reviewed hold key frame when it has one, otherwise under their own frame. This release ships them with review: {"reviewed": false} (physical checks only; the judge + critic review is pending) and review_accept: null in keyframes_index.json; filter them out for reviewed data. More brush pre-grasps will follow (the extraction covered about 60 % of the brush takes).
frame.jpg (in hold seed and pre-grasp folders): the TACO marker-removed camera frame at that moment, cropped around the hand (frame_source: triplet, sequence, camera, raw frame, crop). 9 takes have no marker-removed video and no frame.
novel_grip: true marks holds the reviewers flagged as unusual but sound (index along the handle, fingers on the back of the head, palm-up holds). Grasp types are rule-based Feix names; most hammer / brush power wraps come out as Prismatic 4 Finger or Quadpod (palm just over the 3 mm threshold), the least reliable labels in these cards.
Templates: templates/<use>.npz pools the hold grasps only (pre-grasps excluded).

Load a key frame

import json
from pathlib import Path
import numpy as np
import trimesh

root = Path("NBtools")
index = json.loads((root / "keyframes_index.json").read_text())
verified = [r for r in index if r["status"].startswith("verified")]
kf = root / verified[0]["path"]

info = json.loads((kf / "affordance.json").read_text())
a = np.load(kf / "affordance.npz")
tool = trimesh.load(root / info["card"] / info["mesh"], process=False)   # same vertex order as the arrays

touch = a["contact_mm"] <= 5                       # tool vertices the hand touches
thumb = tool.vertices[a["digit"] == 1]             # where the thumb is
hand = trimesh.Trimesh(a["hand_vertices"], trimesh.load(kf / "hand.obj", process=False).faces, process=False)
print(info["use"], info["status"], int(touch.sum()), "contact vertices,", len(thumb), "thumb vertices")

Use a synthesis template

<card>/[models/<model>/]templates/<use>.npz pools all key frames of one use into one template:

t = np.load(root / "035_power_drill/models/dewalt_dcd799/templates/drilling.npz")
t["p_digit"]      # (n_vertices, 6) soft contact probability per digit (palm, thumb, index, middle, ring, little)
t["contact"]      # (n_vertices,) any digit
t["constraint"]   # 2 must (the actuator digit, e.g. index on the trigger), 1 may, 0 none, -1 avoid
t["must_digit"]   # the digit the 'must' vertices belong to
t["provenance"]   # bits: 1 seen in video, 2 inferred hand, 4 VLM prior, 8 derived
t["instances"]    # (n_grasps, n_vertices) one coherent grasp per row: digit per vertex, 255 none

<use>.json holds the use text, the prior (grasp type, approach, must-touch, avoid), the verified frames, physics (mass, COM) and <use>_check.json the force-closure check per instance.

Other files per tool / model

path	content
`mesh/tool.obj`	the mesh, metres
`mesh/convex/part_NN_<name>.obj`	convex hull per part, for collision
`mesh/labels_v2.npz`, `mesh/affordance_v2.json`, `mesh/faces.json`	per-vertex parts, hand/tool roles, named faces (part - side)
`physics.json`	published mass, centre of mass, inertia
`family/variants.npz`	`V` (n_variants x n_vertices x 3): shape variants with the same vertex ids, so every per-vertex map transfers unchanged
`models/<model>/xray_raw/`	raw x-ray fits of the verified drill frames: camera pose + intrinsics, MANO pose, review overlay

Raw video

Each clip behind a key frame sits next to its key frames, once per folder: <card>/keyframes/<tool model or card>/<clip>.mp4 (web clips: the YouTube source segment t0..t1, native frame rate, up to 720p; DexYCB: the take at 30 fps). A key frame's affordance.json["video"] gives file (relative to the key frame folder, ../<clip>.mp4), keyframe_time_in_video_s, the YouTube url and source_time_s. videos/clips.json lists every clip: url, video id, title, t0/t1 (seconds in the source), fps, frame size, cards, and where its file lives.

Sources and licence

YCB power drill and DexYCB (CC BY-NC 4.0, hence this licence). Spoon / fork / hammer / brush motion capture: TACO (Liu et al., CVPR 2024, CC BY 4.0) and OakInk2 (Zhan et al., CVPR 2024, CC BY-SA 4.0). The spoon / fork key frames whose source is OakInk2 (clips oak_*), and their variants, are derived from OakInk2 and are shared under CC BY-SA 4.0. Please cite TACO and OakInk2 when you use those grasps. YouTube clips are used for research only: frames and overlays are shown for reference, and the videos belong to their owners. Drill and glue model meshes are built from the makers' product photos at the published sizes. Priors come from Qwen3.6-27B. Research use only.

Downloads last month: 58

Total file size:

6.5 GB