FORWARD: Dataset of a forwarder operating in rough terrain

AI summary

Overview

Research area: Field robotics and forestry automation — specifically multimodal sensor datasets for heavy articulated off-road vehicles, terrain traversability, perception, and autonomous control.

Technical level: Intermediate. The paper is a dataset description rather than a methods paper, but it assumes familiarity with sensor modalities (GNSS-RTK, IMUs, LiDAR, CAN-bus logging) and forestry terminology (cut-to-length harvesting, forwarders, StanForD).

Scope (one sentence): The paper documents FORWARD, a roughly 1.1 TB open, synchronized, multimodal dataset of a Komatsu cut-to-length forwarder operating on two Swedish harvest sites, comprising vehicle telematics, high-resolution terrain scans, 360° video, vibration data, production logs, and annotated work elements.

What This Paper Is About

Traversability models for off-road vehicles depend heavily on the vehicle's size, kinematic structure, and driveline, yet the authors state that no publicly available datasets exist for articulated heavy off-road vehicles such as forest machines. This gap, they argue, hinders the application of artificial intelligence and simulation toward higher-precision, more sustainable forestry. The goal of the paper is therefore to publish and document FORWARD — a curated, synchronized, multimodal dataset captured from a real forwarder during both regular wood extraction and deliberately structured experiments, together with analysis scripts and example use cases.

Key Contributions

1. A new open dataset for a heavy articulated off-road vehicle. FORWARD contains about 1.1 TB of data from a large Komatsu forwarder operating at two harvest sites in Ånge municipality, Sweden: Märrviken (data gathered October 2023) and Björsjö (September 2024). The authors state this fills a gap, since no publicly available datasets for articulated heavy off-road vehicles such as forest machines existed to their knowledge.

2. High-resolution terrain coupled to centimeter-accurate machine motion. The sites were laser-scanned with a Riegl Vux 120 sensor from a helicopter (target height 150 m) at approximately 1,500 points per square meter, plus drone photogrammetry, producing point clouds, elevation models, and orthomosaics that can be correlated with RTK-GNSS position and multiple IMU streams.

3. Structured experiment scenarios with systematically varied configurations. Scenarios repeat the same routes with and without steel tracks, with different load weights (0, 10, and 20 tonnes), and at different target driving speeds expressed as percent of maximum speed in terrain gear, on gravel roads, flat quarry surfaces, strip roads, and rough terrain.

4. Annotated work elements and reusable analysis tooling. About 18 hours of regular wood extraction work over three days is annotated from 360°-video into individual work elements, and Python analysis scripts (plus R-scripts and tools such as PDAL and CloudCompare for terrain preprocessing) are published alongside the data. All data and scripts are made long-term publicly available through the Swedish National Data Service (doi:10.71540/89rs-s553).

Main Findings

• Dataset scale and content. The complete dataset comprises about 1.1 TB. Time-series machine data spans 110 h, including GNSS position and speed (0.1 m at 5 Hz), pose (0.1 deg at 5 Hz), acceleration (5 Hz), fuel consumption (5 Hz), operator-set inch speed, and crane slew angle (5 Hz). Codebase additions include wheel-mounted external IMUs (5 Hz, 17 h), operator-seat vibration (6,000 Hz, 27 h), 360°-video (5,376 × 2,688 pixels at 30 fps, 20 h / 560 GB), standard video (3,840 × 2,160 pixels at 60 fps, 373 GB), airborne laser scan (1,500 pt/m², 15 ha / 70 GB), and drone photogrammetry (4,000 pt/m², 15 ha / 56 GB).

• Site differences and sensor availability. Not all modalities exist for both sites — additional sensors were added for the second field trial at Björsjö. No high-frequency vibration data is available from Märrviken. StanForD forwarder production data (.fpr) covers 40 h at Björsjö, while harvester production data (.hpr) covers 110 h and is available for both sites.

• Annotated work element distribution. Over the 18.3 h of video-annotated work at Björsjö: Loading 7.5 h, Unloading 3.3 h, Driving while loading 2.3 h, Driving empty 1.8 h, Driving loaded 1.4 h, Short delays 1.1 h, Other time 0.9 h, and Driving while unloading 0.05 h.

• Fuel consumption modeling is only partly explained by simple variables. A linear regression using driving speed, pitch angle, and booleans for full load and tracks on, fitted to 57,802 observations (about 3 hours of driving), achieved an R² of 0.62 and an RMSE of 4.87 l/h. The authors note this suggests fuel consumption models for rough terrain need more factors than speed, slope, and load weight, and that in this dataset the use of steel tracks appears to affect fuel consumption more than load weight.

• Operator whole-body vibration levels. For a stitched sequence of 11 hours and 20 minutes of regular forwarding work, the vibration dose value (VDV) is 20.0 m/s^1.75 and the daily vibration exposure A(8) is 0.66 m/s², computed per ISO 2631-1; the daily equivalent static compression dose S_ed (ISO 2631-5) is 0.43671 for that sequence. Individual scenarios show widely varying values, for example VDV of 6.36 m/s^1.75 and A(8) of 0.18 m/s² for a gravel, tracks-off, empty, inch_030 away run, versus VDV 59.22 m/s^1.75 and A(8) of 1.04 m/s² for the terrain test circuit with full load at inch_040.

• Synchronization accuracy. Machine data sampled at 5 Hz (GNSS and CAN signals) is the primary temporal reference, and all other sensors were aligned and resampled to it. Because synchronization of video, vibration, and IMUs relied on timestamps with 1 second resolution and in some cases manual identification of synchronization events, the uncertainty is estimated to be on the order of 1 second. Sub-second relationships between high-frequency signals should be interpreted with caution.

• StanForD position resolution is coarser. Machine positions in the StanForD production data are recorded at a coarser time resolution than the 5 Hz used in the other data, so they are upsampled by repeating identical positions until the next position is detected; the difference in time resolution is noticeable when plotted on a map.

• Obstacle-related vibration signatures. In the vertical direction, peaks in acceleration are observed when the forwarder drives over obstacles such as stones and stumps; the dataset includes a clear example of a peak when the forwarder drives down from a 90 cm high stone on the terrain test circuit.

Methodology in Plain English

The researchers instrumented a Komatsu forwarder delivered new in 2023 and ran two field campaigns at two harvesting sites. Manufacturer-installed telematics (RTK-GNSS for position, speed, and heading; CAN-bus engine and driveline signals; crane slew signals) provided the baseline at 5 Hz. They added a 360° camera on the cabin roof, vibration sensors in the operator seat, and multiple IMUs on the wheels. The sites were scanned before harvesting with helicopter-borne LiDAR, and photographed afterwards with drones.

On top of regular production work, they ran controlled experiments. On an old quarry they created a smooth flat baseline; on gravel roads and strip roads they tested slopes of about 5° and 13°; on rough paths with boulders they drove repeated short cycles. In each case they systematically varied load weight, target speed (as a percentage of full speed in terrain gear, set by the operator), and whether steel tracks were mounted. The scenario names encode all of this: location, surface, configuration, and free-text description, for example BRJ_terrain_tracks_on_load_full_inch_020-test_circuit-part_1, where inch_020 means 20% of max speed on low gear (about 2.1 m/s).

The machine's 5 Hz data stream was used as the master clock, and everything else was aligned to it. Video was synced via metadata timestamps, IMUs via deliberately induced characteristic events (rotation about a single axis, brief free-fall), and the high-frequency 6,000 Hz vibration data was aggregated into mean, minimum, maximum, and standard deviation over each 5 Hz interval. Separately, one PhD student watched the 360° video of three days of regular work and annotated it into eight work elements following an established ontology; the authors note that no formal validation of that annotation was performed.

Terrain point clouds were classified and interpolated into elevation models using a more manual approach with R-scripts and existing tools such as PDAL and CloudCompare. Python scripts (mostly NumPy and Matplotlib, plus format-specific libraries) provide a modular processing pipeline covering loading, synchronization, conversion, scenario extraction, saving to npz, plotting, and basic analysis.

Why This Matters

Impact on research. The dataset lets researchers move beyond the forestry standard of judging traversability only by local slope, surface roughness, and bearing capacity. It enables validation and calibration of multibody dynamics simulation models, extraction of automation scenarios and benchmarking data for testing automation solutions, annotation of terrain objects (stones, stumps, logs, vegetation) in images and 3D point clouds, and study of whole-body vibration exposure under different terrain conditions.

Real-world applications.

• Terrain-aware route planning that accounts for how local surface topography affects fuel consumption, effective speed, and vibrations.
• Development of perception and autonomous control systems for forest machines, including obstacle avoidance and loading/unloading of logs with cranes.
• Automatic work element detection (AWED) in forwarding, using annotated work elements together with machine, position, and terrain data.
• Simulation development, including auto-generation and calibration of forestry machine simulators and synthetic data generation to complement physical testing.
• Occupational health analysis, using the measured vibration dose values and daily vibration exposures to assess operator exposure across terrain and configurations.

Industry relevance. The work involves Komatsu Forest AB and Skogforsk, the Swedish Forest Research Institute, alongside Umeå University and the Swedish University of Agricultural Sciences. The stated aim is higher levels of automation and precision forestry, which the authors link to achieving sustainability goals. The paper also notes the value of statistics on machine utilization in relation to terrain properties for the forestry industry.

Future Directions

• Refining synchronization. Because the estimated uncertainty is on the order of 1 second, the authors state that sub-second relationships between high-frequency signals such as IMU and vibration measurements should be interpreted with caution unless further refinement of the synchronization is performed.
• Better fuel consumption models. The linear regression with speed, pitch, and load/track booleans reached only R² 0.62 and RMSE 4.87 l/h, indicating that additional factors beyond speed, slope, and load weight are needed for rough-terrain fuel modeling.
• Automatic work element detection. The authors see opportunities to advance automatic work element detection in forwarding work by combining annotated work elements with machine, position, and terrain data.
• Validation of annotations and simulator calibration. The work element annotation received no formal validation; any user can validate it using the annotation file and the lightweight 360° videos. Separately, the data is intended to support auto-generation and calibration of forestry machine simulators and automation scenario descriptions.

Target Audience

Researchers and engineers working on field robotics, off-road vehicle traversability, forestry automation, and physics-based simulation. It is also relevant to machine learning practitioners needing real-world multimodal sensor data with terrain ground truth, to developers building or calibrating forestry machine simulators, and to occupational health researchers studying whole-body vibration exposure in heavy machinery. The involvement of Komatsu Forest AB and Skogforsk makes it directly relevant to forestry industry R&D groups. Readers looking for a benchmark-model results paper will not find one here — this is a dataset and data-description paper.

Authors’ abstract

We present FORWARD, a high-resolution multimodal dataset of a cut-to-length forwarder operating in rough terrain on two harvest sites in the middle part of Sweden. The forwarder is a large Komatsu model equipped with vehicle telematics sensors, including global positioning via satellite navigation, movement sensors, accelerometers, and engine sensors. The forwarder was additionally equipped with cameras, operator vibration sensors, and multiple IMUs. The data includes event time logs recorded at 5 Hz of driving speed, fuel consumption, machine position with centimeter accuracy, and crane use while the forwarder operates in forest areas, aerially laser-scanned with a resolution of around 1500 points per square meter. Production log files (Stanford standard) with time-stamped machine events, extensive video material, and terrain data in various formats are included as well. About 18 hours of regular wood extraction work during three days is annotated from 360-video material into individual work elements and included in the dataset. We also include scenario specifications of conducted experiments on forest roads and in terrain. Scenarios include repeatedly driving the same routes with and without steel tracks, different load weights, and different target driving speeds. The dataset is intended for developing models and algorithms for trafficability, perception, and autonomous control of forest machines using artificial intelligence, simulation, and experiments on physical testbeds. In part, we focus on forwarders traversing terrain, avoiding or handling obstacles, and loading or unloading logs, with consideration for efficiency, fuel consumption, safety, and environmental impact. Other benefits of the open dataset include the ability to explore auto-generation and calibration of forestry machine simulators and automation scenario descriptions using the data recorded in the field.