Machines that move
Train driving, robot and civil drone models
Auto-label your camera, lidar and radar logs and test every release candidate against the last release.
What it's worth
Cut the hand-labelling behind every release
- to label one 25-second, 10 Hz lidar scene by hand
- 10+ hours
- to hand-label one lidar frame of 20–60 objects at about USD 1–2 a box
- USD 20–120
- lower cost per 3D box with neural auto-labelling, vendor-reported by Bosch
- Up to 70%
Sources: Auto4D, Uber ATG and University of Toronto, 2021 (arXiv 2101.06586); Shaip and Label Your Data on objects per frame; an informal industry write-up on 3D-box prices, arithmetic ours; Bosch Engineering, vendor figure.
Use cases
Rare cases differ for every fleet
| Use case | Raw data | Rare cases to find | Adapt per | Release test |
|---|---|---|---|---|
| Autonomous driving | Cameras, lidar, radar, GNSS/IMU | Hard braking, unusual road users, poor visibility | Market, season | Distance to collision in simulation |
| Robots and civil drones | Cameras, lidar, IMU | Near-misses, unusual obstacles | Robot model, site | Collisions in simulation |
| Trucks and logistics yards | Cameras, lidar, radar, trailer telemetry | Dock reversing, people between trailers, poor night light | Yard layout, trailer type | Missed people in reversing scenarios in simulation |
| Mines and ports | Vehicle cameras, lidar, telemetry | Dust, blind spots, people in vehicle lanes | Site, machine | Missed hazards in simulation |
Autonomous driving
- Raw data
- Cameras, lidar, radar, GNSS/IMU
- Rare cases to find
- Hard braking, unusual road users, poor visibility
- Adapt per
- Market, season
- Release test
- Distance to collision in simulation
Robots and civil drones
- Raw data
- Cameras, lidar, IMU
- Rare cases to find
- Near-misses, unusual obstacles
- Adapt per
- Robot model, site
- Release test
- Collisions in simulation
Trucks and logistics yards
- Raw data
- Cameras, lidar, radar, trailer telemetry
- Rare cases to find
- Dock reversing, people between trailers, poor night light
- Adapt per
- Yard layout, trailer type
- Release test
- Missed people in reversing scenarios in simulation
Mines and ports
- Raw data
- Vehicle cameras, lidar, telemetry
- Rare cases to find
- Dust, blind spots, people in vehicle lanes
- Adapt per
- Site, machine
- Release test
- Missed hazards in simulation
How it works
Raw sensor logs in. Tested releases out.
Melt dataPrepare the data · 01–03
01Ingest
Load camera, lidar, radar and GNSS/IMU logs, drop corrupt files, catalogue by site, weather and time of day, and anonymise faces and plates.
02Mine
Find near-misses, unusual obstacles and poor visibility, and drop near-identical routine runs.
03Label
Auto-label 3D boxes, segmentation and trajectories with large offboard models inside the platform; your engineers correct only labels below the confidence they set.
Forge intelligenceTrain and test the model · 04–06
04Train
Retrain your shared base model on the full, versioned dataset when new site or machine logs land; every release traces to its exact data.
05Adapt
Fine-tune the base model for each machine, site or market, and tune it again as seasons and traffic change.
06Validate
Optimise each release candidate for your onboard computer, such as NVIDIA DRIVE or Jetson, then run scenario tests and closed-loop simulation against the last release.
What it trains
From perception to on-device models
Perception
Detect 3D objects, lanes, signs and obstacles, and predict occupancy from multi-camera input.
Sensor fusion
Merge surround cameras, lidar, radar, IMU and GNSS into one bird's-eye view.
Prediction and planning
Predict how people and vehicles will move, and plan motion on roads, in warehouses and on work sites.
End-to-end and world models
Train transformers that map sensor data straight to a control policy, and train world models.
Simulation and synthetic data
Generate rare, safety-critical scenarios with world models. Run log-replay and object-level closed-loop tests on the training GPUs.
On-device optimisation
Distil and quantise models for onboard computers.
Adapt per site
Train for the places your machines run
Layouts, weather and traffic change from site to site and through the year.
Adaptation runs on your own logs from each site, and no customer's data is mixed with another's.
Layouts and rules
Lane markings, signs, junction types, site layouts, left- and right-hand traffic.
Weather and light
Heat and dust, rain and fog, snow and ice, working after dark.
Traffic
Dense traffic, pedestrians, two-wheelers, trucks and uncontrolled junctions.
Quality
Regressions stop at the gate
Every release candidate runs the last release's test sets; a drop beyond your margin blocks it until you sign off.
Your team sets the metric and margin for each test set, such as detection recall, false-positive rate and distance to collision in simulation, each with its 95% confidence interval.
Machines that move
release v13 → release candidate v14
| Test set | Cases | Last release | Release candidate | Margin | Result |
|---|---|---|---|---|---|
| Night pedestrians, detection recall | 2,140 | 0.94 | 0.95 | −0.02 | Pass |
| Dust and haze, detection recall | 860 | 0.81 | 0.80 | −0.02 | Within margin |
| Occluded cyclists, detection recall | 1,920 | 0.90 | 0.86 | −0.02 | Fail −0.04 |
| Reversing at docks, people behind trailer, detection recall | 640 | 0.86 | 0.90 | −0.02 | Pass |
| Simulation, min distance to collision | 5,000 episodes | 1.8 m | 2.1 m | −0.2 m | Pass |
Night pedestrians, detection recall
- Cases
- 2,140
- Last release
- 0.94
- Release candidate
- 0.95
- Margin
- −0.02
- Result
- Pass
Dust and haze, detection recall
- Cases
- 860
- Last release
- 0.81
- Release candidate
- 0.80
- Margin
- −0.02
- Result
- Within margin
Occluded cyclists, detection recall
- Cases
- 1,920
- Last release
- 0.90
- Release candidate
- 0.86
- Margin
- −0.02
- Result
- Fail −0.04
Reversing at docks, people behind trailer, detection recall
- Cases
- 640
- Last release
- 0.86
- Release candidate
- 0.90
- Margin
- −0.02
- Result
- Pass
Simulation, min distance to collision
- Cases
- 5,000 episodes
- Last release
- 1.8 m
- Release candidate
- 2.1 m
- Margin
- −0.2 m
- Result
- Pass
Release candidate blocked: 1 test set beyond its margin
Interface preview · sample data
Your data
Your logs and weights stay yours
- Ownership
- Your data, labels and trained weights stay yours. Bytemelt has no access to them: your contract says so, and the platform is built so only accounts your team controls can open them.
- Anonymisation
- Your team anonymises faces and plates at ingest with the platform's tools, using rules you set, or before the files leave your own systems.
- Location
- Your data is stored and processed only at the partner data centre your contract names, in the Middle East and outside the EU. The contract names the site and country before any data moves, and data leaves the site only when your team moves it. Personal data from the EU moves only under transfer safeguards such as Standard Contractual Clauses, and we give your data-protection officer the site details a transfer impact assessment needs.
Questions autonomy teams ask
How do we load a large backlog of logs?
We size storage for your backlog and what your fleet adds each month, and your contract names the transfer route and loading schedule. Your team runs every transfer into its own workspace.
How does a model get to the onboard computer?
Stage 06, above, optimises and tests each release candidate for it. Your team exports the weights and deploys them.
Size compute for your fleet logs
In one hour, the founder maps your sensor logs and models to the six stages and sizes the compute.
contact@bytemelt.ai