dhyuti-n/feels-cup-dynamics

feels — Cup-Dynamics Proprioceptive Telemetry

What's in the cup? — Can a robot arm tell what's inside a sealed, opaque coffee cup without looking inside, purely from how the contents feel when swirled?

This dataset is the joint-torque telemetry behind the feels project. A 6-DOF UFactory xArm850 (via the Viam platform) picks a lidded cup off a taped table square, lifts it to travel height, executes a horizontal circular swirl, and places it back. Throughout the macro it logs TCP pose, joint angles, and per-joint actuator torques at ~4–5 Hz. The swirl excites the contents — water sloshes and lags, loose candy/screws chatter, a rigid mass just adds inertia, an empty cup does nothing — so the torque signature encodes the payload.

Task

Given one recording's telemetry (especially the SWIRL phase), predict the cup's contents. 9 labeled classes spanning loadedness (empty → full) and rigidity (liquid ↔ loose grains ↔ rigid solid):

empty_cup, water_025pct, water_050pct, water_075pct, water_100pct, solid_020g, solid_050g, solid_100g, screws_100g.

The interesting cases are same mass, different contents (e.g. a 100 g rigid solid vs. 100 g of loose screws) — mass alone can't separate them; the dynamics have to.

Structure

• One example per recording (trial). 106 trials total.
  • train: 85 · test: 21
• Split is stratified by (label, motion_path) at the recording level (seed 42, ~80/20) so rows within a trial never leak across the split.

Fields

field	type	description
file	string	source recording filename
label	string	ground-truth contents (one of the 9 classes)
motion_path	string	swirl (orbital agitation present) or normal (pick/place only)
split	string	train / test (redundant with the split it lives in)
n_rows	int32	number of telemetry samples in the trial
elapsed_sec	list[float32]	timestamps (seconds from macro start)
tcp_x_mm,tcp_y_mm,tcp_z_mm	list[float32]	tool-center-point position
tcp_ox,tcp_oy,tcp_oz,tcp_theta_deg	list[float32]	TCP orientation (axis-angle)
q1_deg…q6_deg	list[float32]	joint angles
tau1_nm…tau6_nm	list[float32]	per-joint actuator torques (the payload signal)
gripper_pos	list[float32]	gripper position
gripper_holding	list[bool]	gripper holding flag
phase	list[string]	per-sample phase tag

Every list column is the time series for that trial and has length n_rows.

Phases (phase)

INIT → REACH_GRASP → APPROACH → GRIPPER_CLOSE → LIFT_TO_TRAVEL → TRAVEL_SETTLING → TRAVEL_STABLE_HOLD → SWIRL → DESCEND_PLACE → GRIPPER_OPEN → BACK_OFF → RETRACT_TRAVEL → RETURN_HOME

The discriminative window is SWIRL (and the surrounding travel phases).

Usage

from datasets import load_dataset

ds = load_dataset("dhyuti-n/feels-cup-dynamics")
ex = ds["train"][0]
print(ex["label"], ex["motion_path"], ex["n_rows"])

# torque time series for joint 2 during the swirl
import numpy as np
phase = np.array(ex["phase"])
tau2  = np.array(ex["tau2_nm"])
swirl_tau2 = tau2[phase == "SWIRL"]

Honest caveat (read before trusting a model)

The project's own baseline analysis found the static hold carries no reliable payload signal — absolute hold torques are identical across contents to 3 decimals, and the tare baseline is contaminated by motion drift (~10 N·m noise vs. a ≤2 N·m payload signal). Discrimination has to come from the dynamics of the SWIRL phase, not static torques. Treat the normal (non-swirl) recordings as negative/control examples, not as classifiable payload data. Sampling at ~4–5 Hz also cannot resolve the 5–25 Hz rattle band.

Collection

• Robot: UFactory xArm850 (6-DOF) + parallel-jaw gripper, orchestrated via Viam.
• Date: 2026-10-29 (SF Hacks hackathon).
• Protocol: fixed pick → swirl → place macro from a calibrated table square. The embedding encodes the motion as much as the contents, so the protocol is held fixed across trials.

License

MIT.