A globally known AI compute leader and an autonomous driving developer
L4 perception annotation at 99.9% accuracy across LiDAR, radar, camera and driver-monitoring data
Published 24 July 2026
At a glance
| Industry | Autonomous driving | an AI compute platform leader and an AV developer |
|---|---|---|
| Region | Global clients | delivery from dedicated AV centres in Malaysia and Indonesia |
| Modalities | 4 | LiDAR point clouds · multi-camera detection · radar fusion · driver-monitoring data |
| Task types | 5 | object detection · scene segmentation · 3D point-cloud annotation · behaviour prediction · sensor fusion |
| Accuracy benchmark | 99.9% | on L4-level safety-critical scenarios |
| Standard SLA | 95%+ | Lifewood-wide accuracy standard; this programme runs above it |
| Review process | Dual-layer human-in-the-loop | first-pass annotator plus independent second-pass reviewer |
| Audit record | Per asset | timestamped approval record retained for post-delivery audit |
| Status | Active partnership | supporting both clients' autonomous driving programmes |
This is a Lifewood Data Technology case study in Autonomous driving — an engagement delivered for a globally known AI compute leader and an autonomous driving developer across Global · Asia-Pacific. L4 perception annotation at 99.9% accuracy across LiDAR, radar, camera and driver-monitoring data
L4 autonomy leaves no room for a statistical error budget
L4 autonomy programmes require multi-modal sensor data annotated to safety-critical accuracy: LiDAR point clouds, multi-camera detection, radar fusion and driver-monitoring system data, with throughput, accuracy and audit trail all non-negotiable at once. The reason the accuracy bar sits where it does is that the error budget is physical rather than statistical. A mislabelled pedestrian in training data is not a percentage point in a report; it is a failure mode in a vehicle. That changes what a data partner is being asked for. It is not a labelling service with a quality score attached — it is an evidence chain, where every asset needs to be traceable to who annotated it, who reviewed it, and when it was approved, long after delivery.
Four sensor modalities annotated to one standard, in dedicated AV centres
Lifewood operates as an autonomous-driving data partner for both clients, supplying driver-monitoring system data and supporting autonomous-driving AI model training. The programme spans four modalities that have to agree with each other: LiDAR point clouds give 3D structure, multi-camera detection gives semantics, radar adds velocity and holds up in weather that defeats cameras, and driver-monitoring data covers the cabin. Sensor fusion — reconciling all four into one consistent scene — is the part where single-modality vendors usually stop. 1. Dedicated AV centres, not shared capacity. Delivery runs through dedicated autonomous-vehicle centres in Malaysia and Indonesia inside Lifewood's 40+ centre footprint. Dedicated rather than shared, because AV annotation tooling, training and security review differ enough from general labelling that mixing them degrades both. 2. A trained bench rather than surge headcount. Throughput comes from trained, dedicated teams rather than from relaxing review. 414,120 training hours were delivered across the Bangladesh workforce during 2025, an average of 60 hours per person there. 3. Dual-layer human-in-the-loop review. A first-pass annotator and an independent second-pass reviewer work against a customer-approved gold set, with a 95%+ inter-annotator agreement threshold. 4. A per-asset audit record. Every asset keeps a timestamped approval record, so a programme can be audited long after delivery. Throughput and safety-critical accuracy are treated as one requirement rather than a trade-off, which is the only way both survive contact with a delivery schedule.
Annotation accuracy holds at 99.9% on L4-level scenarios
Accuracy is benchmarked at 99.9% for L4-level safety-critical scenarios, against the 95%+ standard that applies across every Lifewood programme. Inter-annotator agreement is tracked as a separate number at a 95%+ threshold, because per-item accuracy alone describes a vendor's agreement with itself, while agreement between two independent reviewers describes whether the specification is genuinely shared.
Related Lifewood services
Verified outcomes
| Metric | Value | Baseline | How measured |
|---|---|---|---|
| Annotation accuracy | 99.9% | vs the 95%+ standard Lifewood SLA | Benchmarked on L4-level safety-critical scenarios |
| Inter-annotator agreement | 95%+ | threshold, not a period average | Independent second-pass reviewer vs first-pass annotator |
| Modalities covered | 4 | single-modality vendors typically stop before fusion | LiDAR, camera, radar, driver-monitoring |
Method and verification. Figures on this page are Lifewood-reported. Annotation accuracy is measured against a customer-approved gold set under the dual-layer human-in-the-loop review described above, with inter-annotator agreement tracked separately as a second number. Every batch carries a timestamped approval record available for client and procurement audit. The 414,120 training-hours figure covers the Bangladesh workforce in 2025 and is a workforce investment rather than an output of this engagement. Both clients are unnamed on this page under the confidentiality terms of their agreements.
Questions about this programme
On Lifewood's autonomous driving annotation programmes accuracy is benchmarked at 99.9% for L4-level scenarios, against a 95%+ standard SLA elsewhere.
A Lifewood AV perception programme covers four modalities — LiDAR point clouds, multi-camera detection, radar fusion and driver-monitoring data — and they have to agree with each other.
Lifewood delivers autonomous driving annotation through dedicated AV centres in Malaysia and Indonesia, inside a 40+ delivery centre footprint.
On AV annotation programmes Lifewood treats throughput and safety-critical accuracy as one requirement rather than a trade-off.
Run a similar program with Lifewood?
Tell us your scope and accuracy bar. We will scope a comparable engagement within one call.
Talk to our team