Dataset card

README.md

@@ -1,61 +1,131 @@
-# FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models.
+---
+license: mit
+pretty_name: FactoryNet
+size_categories:
+  - 100M<n<1B
+task_categories:
+  - time-series-forecasting
+  - tabular-classification
+tags:
+  - industrial
+  - robotics
+  - anomaly-detection
+  - time-series
+  - sim-to-real
+  - manufacturing
+  - predictive-maintenance
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: data/*.parquet
+---
 
-FactoryNet is a large-scale, unified dataset designed to advance Industrial Time-Series Foundation Models within robotic manufacturing. By providing highly standardized, high-frequency telemetry across different robot morphologies (UR3, KUKA), specialized end-effectors (screwdrivers, grippers), and tens of thousands of simulated counterfactuals, FactoryNet serves as a rigorous benchmark for sequence modeling, anomaly detection, and sim-to-real gap analysis.
+# FactoryNet
 
-This repository is completely anonymized for double-blind review.
+A multi-embodiment industrial time-series corpus: **56,591 end-to-end task executions**
+(14,847 real, 41,744 simulated), **113M logged timesteps**, **7 embodiments**, **4 tasks**,
+**27 annotated anomaly types** with healthy baselines and counterfactual pairs.
 
----
+Every signal from every source is mapped into one control-theoretic schema —
+**Setpoint, Effort, Feedback, Context (S-E-F-C)** — so a single dataloader works across a
+6-DOF arm and a 4-axis CNC gantry, and *commanded versus realized* dynamics are readable
+as explicit prediction residuals rather than opaque reconstruction scores.
 
-## Repository Structure
+*This repository is anonymized for double-blind review.*
 
-The repository is structured into two main components: the unified structured data (`/data`) and the Sim2Real gap evaluation artifacts (`/artifacts`).
+## Quick start
 
-* **`data/`**: Contains the massive-scale, chunked Parquet files adhering to the S-E-F-C schema.
-  * `aursad_*.parquet`, `cnc_*.parquet`, `voraus_*.parquet`: Pre-existing structured datasets.
-  * `factorywave_*.parquet`: Real-world UR3 and KUKA telemetry (including specialized screwdriver torque data and episode metadata).
-  * `simulations_baseline_*.parquet`: 10,000+ baseline pick-and-place episodes generated in Isaac Sim 4.5.0.
-  * `simulations_counterfactual_*.parquet`: Matched counterfactual episodes for causal and anomaly modeling.
-* **`artifacts/`**: Contains the targeted Sim2Real benchmarking toolkit.
-  * `real_csv/` & `sim_csv/`: Paired episode CSVs matching real-world executions with Isaac Sim equivalents.
-  * `pick_configs/`: Per-episode JSON configurations driving the simulation parameters.
-  * `gap_reports/` & `summary/`: Aggregated gap analysis, inventories, and per-episode markdown reports.
+```python
+from datasets import load_dataset
 
----
+ds = load_dataset("FactoryNet4/factorynet", split="train", streaming=True)
+print(next(iter(ds)))
+```
 
-## The S-E-F-C Master Schema
+Or read the Parquet directly, which is usually what you want for time series:
 
-To facilitate foundation model training across heterogeneous hardware, all telemetry in the `data/` directory has been mapped to a unified **Setpoint-Effort-Feedback-Context (S-E-F-C)** schema. 
+```python
+import pandas as pd
+from huggingface_hub import hf_hub_download
 
-Rather than dealing with vendor-specific labels, models can train on standardized representations:
-* **Setpoint (`setpoint_*`)**: The commanded states (position, velocity, acceleration, target torque).
-* **Effort (`effort_*`)**: The hardware's physical exertion (current, voltage, force, torque).
-* **Feedback (`feedback_*`)**: The measured physical state (actual position, velocity).
-* **Context (`ctx_*`)**: Metadata and discrete states (anomaly labels, task phases, hardware modes, execution times).
+p = hf_hub_download("FactoryNet4/factorynet", "data/factorywave_ur3_000.parquet",
+                    repo_type="dataset")
+df = pd.read_parquet(p)
 
-*Note: Specialized screwdriver telemetry (e.g., tightening states, measured shank torque) is fully integrated into the tool-specific effort and context columns.*
+# every setpoint/effort pair, for any embodiment, without vendor-specific names
+S = [c for c in df.columns if c.startswith("setpoint_")]
+E = [c for c in df.columns if c.startswith("effort_")]
+```
 
----
+## The S-E-F-C schema
+
+One wide table per source: one row per control tick, joined to an episode table by
+`episode_id`. The column prefix carries the role; the suffix `_i` is the positional joint
+index 0..5 (base → wrist), so the same column means the same slot on every arm. A column a
+source does not expose is written as null rather than dropped, so all sources share one
+schema.
+
+| Role | Prefix | What it is | Examples |
+|---|---|---|---|
+| **Setpoint** | `setpoint_*` | commanded intent | target joint position, velocity, acceleration; target TCP pose; `gripper_command` |
+| **Effort** | `effort_*` | actuation energy expended | motor current, joint torque, commanded torque, estimated contact force |
+| **Feedback** | `feedback_*` | measured physical outcome | encoder position and velocity, TCP pose, tool accelerometer |
+| **Context** | `ctx_*` | environment and static state | payload mass, task phase, fault labels, safety/robot mode, speed scaling, I/O bits |
+
+This separation is the point of the corpus. Existing industrial datasets log sensor
+*outcomes* without distinguishing what the controller asked for from what the machine did,
+which makes cross-machine dynamics learning hard to set up at all.
+
+## Composition
+
+| Source | Machine | Tasks | Faults | Episodes | Timesteps |
+|---|---|---|---|---|---|
+| Lab (real) | UR3 | P&P, Screw, Peg | yes | 8,863 | 11M |
+| Lab (real) | KUKA KR10 | P&P | yes | 1,799 | 4M |
+| Open (real) | voraus-AD (Yu-Cobot) | P&P | yes | 2,122 | 16M |
+| Open (real) | AURSAD (UR3e) | Screw | yes | 2,045 | 3M |
+| Open (real) | UMich CNC | Machining | yes | 18 | 18K |
+| Synthetic | Isaac Sim — UR3, UR5, UR10, UR30, KR10 | P&P | yes | 41,744 | 79M |
+| **Total** | | | | **56,591** | **113M** |
+
+1,553 real counterfactual episodes accompany the faulty runs. Of the 10,662 lab episodes,
+~28% are healthy and ~72% contain an injected fault.
+
+### Synthetic track
+
+41,744 validated Pick & Place episodes across five arms, generated in NVIDIA Isaac Sim 5.1
+/ Isaac Lab 2.3 (GPU-batched PhysX at 500 Hz; control and logging at 125 Hz for UR and
+83.3 Hz for the KR10, no resampling). Six fault classes: payload addition, motor
+miscommutation, gripper activation failure, gripper release, collision, path obstacle.
+
+Every episode passes a two-stage validity gate — structural (plausible length, complete
+phases, part placed, injected event present, non-zero contact force) and trace (every
+fault must leave a measurable physical effect). **41,744 of 50,000 generated episodes
+pass.** This guarantees that a fault label corresponds to a real physical effect, but it
+also biases the synthetic faults toward more detectable instances, and acceptance rates
+differ by fault class. UR3 and KR10 motion profiles are fitted to real recordings; UR5,
+UR10 and UR30 have no real counterpart and are scaled from the UR3 fit.
 
-## Sim2Real Gap Artifacts (UR3 Pick-and-Place)
+## Known limitations
 
-In addition to the raw structured data, this repository provides artifacts explicitly generated for kinematic and effort alignment benchmarking between real UR3 executions and Isaac Sim 4.5.0.
+Stated here because they determine what the corpus can be used for:
 
-### Episode Pairing and Chronology
-Real and sim episodes within the `artifacts/` folder are strictly paired by filename (e.g., `real_csv/<uuid>.csv` ↔ `sim_csv/<uuid>.csv`). 
+- **UR5, UR10 and UR30 exist only in simulation.** Any multi-arm result on them is sim-to-sim.
+- **The validity gate favours detectable faults**, so synthetic fault difficulty is not representative.
+- **Recording context partly predicts fault labels** on both real robots (program, day, speed override), so real fault evaluation should stay within a single recording session.
+- **Simulated motor current is not calibrated** to the real robots out of the box. Task and cycle structure transfer from simulation; fault signatures largely do not.
+- The KUKA KR10 (KSS 8.3) does not expose joint velocities, commanded TCP pose, or TCP force/torque over RSI; those channels are null.
 
-Chronological payload buckets were assigned using boundary episodes to validate dynamic alignment:
-* **0.6 kg**: Start → `0ab540a6-7210-4531-b2ed-6a04c7b90ff1`
-* **0.3 kg**: After that → `29312353-3555-4eba-b2fb-3e2495e89782`
-* **1.2 kg**: After that → `640c65a7-2aee-4b58-94de-334f37832e14`
+## Licensing
 
-### Gap Metrics
-For each matched real/sim pair, the provided reports track:
-* **Joint RMSE (deg)**: Error across individual joints.
-* **TCP position RMSE (mm) / EE L2 RMS (mm)**: Pooled and step-wise Euclidean position errors.
-* **TCP rotvec RMSE (mrad)**: Rotational orientation gaps.
-* **W1 effort (A)**: Wasserstein-1 distance on `joint_current_*` to proxy effort alignment.
+Novel laboratory and synthetic data: **MIT**. Adapted open-source subsets
+(voraus-AD, AURSAD, UMich CNC) retain their original licenses and are redistributed under
+them. Cite the original sources when using those subsets.
 
-*(Alignment is achieved by sorting by time, splitting by task phase, and linearly resampling timeframes to [0, 1] prior to calculating distances).*
+## Code
 
-### Scope and Limits
-The Sim2Real artifact package is strictly designed to support **kinematic agreement analysis** and **effort-proxy comparisons**. It is explicitly *out of scope* for contact-accurate dynamics validation, visual-domain realism, or claims of identical actuator torque physics.
+Adapters, simulation configs, the full transfer study and every result file behind the
+paper's figures are in the anonymous code repository linked from the paper. A verification
+script there re-resolves all 126 plotted values against the raw result files with no data
+download required.