Industry Research

Appearance

Dynamic Occupancy Freespace ​

Executive Summary ​

  • Dynamic occupancy freespace models predict both what space is occupied and what space is free now or in the future.
  • The recent research line includes implicit 4D occupancy-flow world models, self-supervised LiDAR occupancy fields, and unified forecasting benchmarks.
  • DIO decomposes dynamic scenes into instance-aware implicit 4D occupancy-flow components.
  • UnO learns unsupervised 4D occupancy fields from LiDAR, using depth and free-space rendering objectives to learn perception and forecasting without manual object labels.
  • UniOcc standardizes occupancy prediction and forecasting across datasets and adds per-voxel flow annotations, making it useful for comparing dynamic occupancy methods.
  • For autonomy, the key value is not only future occupied cells; it is knowing which rays and regions are observed free, unknown, occluded, or expected to become occupied.

Problem Fit ​

  • Use this method family when planning needs explicit free-space and future occupancy rather than only tracked object trajectories.
  • It fits dynamic scenes where objects are partly occluded, not well detected as boxes, or not represented by a fixed object class list.
  • It is useful for safety buffers because occupancy can represent object extent, swept volume, and unmodeled obstacles.
  • It is especially relevant to low-speed autonomy around pedestrians, baggage trains, forklifts, carts, loading equipment, and vehicles that perform non-lane-following maneuvers.
  • It is less suitable as a standalone semantic detector; many methods emphasize geometry and motion more than fine class labeling.
  • It should be paired with uncertainty estimates because future freespace can be safety-critical and wrong in asymmetric ways.

Method Mechanics ​

  • A dynamic occupancy model maintains a spatial representation of occupied, free, and unknown space over time.
  • DIO represents the world as a decomposable implicit 4D occupancy-flow model, using instance prompts to complete current shapes and forecast occupancy-flow evolution.
  • DIO can take prompts from detectors, so it bridges explicit object proposals with dense implicit shape and motion fields.
  • UnO trains from LiDAR self-supervision rather than human object labels, learning an occupancy field that can render depth and free-space evidence along LiDAR rays.
  • UnO's balanced occupancy sampling is important because naive free-space-heavy training can become under-confident about occupied actors.
  • UniOcc provides a unified benchmark for current-frame occupancy prediction and future occupancy forecasting across real-world and simulated datasets.
  • Radar-centric dynamic occupancy grid mapping adapts inverse sensor models, field-of-view handling, and state computation to radar measurements rather than LiDAR assumptions.
  • Practical systems usually need two coupled products: a current freespace layer for immediate collision checking and a future occupancy layer for planning.

Inputs and Outputs ​

  • Input: historical LiDAR sweeps, camera features, radar returns, tracked object prompts, or BEV features depending on the method.
  • Input: ego pose and timestamp history for motion compensation.
  • Input: sensor rays or visibility masks for free-space supervision.
  • Optional input: 3D detector instance prompts, map lanes, route intent, or cooperative sensor data.
  • Output: current occupancy or freespace grid in BEV or 3D.
  • Output: future occupancy over multiple horizons.
  • Output: occupancy flow or scene flow vectors for dynamic cells.
  • Optional output: instance-conditioned occupancy fields, semantic occupancy, visibility confidence, and unknown/occluded masks.

Assumptions ​

  • Sensor pose and timing are accurate enough to align historical observations.
  • The training data contains enough dynamic examples to learn non-trivial motion, not only static completion.
  • Free-space labels from LiDAR rays are reliable; they prove observed emptiness only along visible rays, not behind occluders.
  • Instance prompts from detectors are useful but may be incomplete; DIO-style decomposition depends on prompt quality.
  • Future occupancy is conditioned on recent motion and scene context, but it cannot infer hidden intent with certainty.
  • The planner understands that free, unknown, and future occupied have different safety meanings.

Strengths ​

  • Occupancy captures object extent and non-object hazards better than point tracks alone.
  • Freespace supervision from sensor rays can scale with raw logs and reduce dependence on manual labels.
  • Occupancy flow gives planners a compact dynamic field for swept-volume reasoning.
  • Implicit 4D fields can represent continuous space and time more flexibly than fixed grid rollouts.
  • Decomposable instance prompting helps connect dense occupancy to detector and tracker outputs.
  • Unified benchmarks make cross-method comparison less dependent on private label-generation pipelines.
  • Radar or LiDAR dynamic occupancy grids can run as interpretable safety layers alongside learned models.

Limitations and Failure Modes ​

  • Future freespace can be dangerously overconfident when hidden actors emerge from occlusion.
  • Self-supervised LiDAR objectives inherit LiDAR visibility limits and may learn sensor-specific blind spots.
  • Detector prompts can cause missed-instance holes in instance-conditioned models.
  • Occupancy flow may be smooth and plausible while missing sudden stops, U-turns, or personnel stepping into view.
  • Free-space rendering can encourage conservative emptiness near rays but weak object extent in sparse regions.
  • Dense 3D forecasting can be expensive at long horizon and high resolution.
  • Benchmarks may use road-centric labels and miss indoor, airport, port, or warehouse movement patterns.

Evaluation Notes ​

  • Evaluate current occupancy, current freespace, future occupancy, and flow separately.
  • Include horizon-specific metrics, because a model can be useful at 0.5 s and unreliable at 3 s.
  • Report false-free-space rate near the ego path, not only occupancy IoU.
  • Split results by dynamic/static cells, occluded/visible cells, actor class, range, and speed.
  • Use UniOcc-style unified settings when comparing methods across nuScenes, Waymo, CARLA, and cooperative data.
  • Stress test with missing detector prompts, delayed sensors, dropped LiDAR sweeps, radar ghost returns, and unexpected stopped objects.
  • For airport validation, add pushback, baggage train turns, service-road crossings, stand-entry maneuvers, workers behind vehicles, and aircraft occlusions.

AV and Indoor/Outdoor Relevance ​

  • On-road AVs: useful for dense occupancy forecasting around lane changes, turns, occlusions, and unusual objects.
  • Airport AVs: very high relevance because movement is often low-speed, non-lane-bound, and filled with irregular equipment.
  • Indoor robots: useful for forklifts, humans, doors, carts, shelves, and blind corners when RGB-D or LiDAR rays provide freespace evidence.
  • Outdoor yards and ports: useful for containers, trailers, cranes, debris, and mixed human-machine motion.
  • Warehouse robots may prioritize 2D/2.5D dynamic occupancy, while airport and road AVs often need 3D clearance around overhangs and aircraft.
  • The same representation can support ODD management: slow down when predicted freespace confidence drops or unknown space grows.

Implementation/Validation Checklist ​

  • Define the difference between free, unknown, occupied, occluded, and future occupied before training labels are generated.
  • Preserve sensor ray geometry so free-space labels remain auditable.
  • Build a baseline dynamic occupancy grid or raycasting model before adding implicit neural models.
  • Validate ego-motion compensation using static-scene replay; drift will appear as false flow.
  • Track calibration and time sync as model inputs or gating signals.
  • Evaluate false freespace inside braking distance and swept-path envelopes.
  • Include a prompt-drop benchmark if using DIO-style detector-conditioned decomposition.
  • Include a raw-log self-supervised training path if following UnO-style LiDAR occupancy fields.
  • For airport use, define conservative planner behavior for unknown voxels under aircraft wings, near terminal glass, and around parked GSE.

Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim