PaperlayerView exact revision on arXiv

Source-linked AI summary

Rendering-Free Event Frame Simulation for Fast Reinforcement Learning of Autonomous Robot Navigation

Dequan Ou, Austin Lopez, Quentin Missinne, Christophe De Wagter, Guido C. H. E. de Croon

arXiv:2610.37543v1cs.RO

TL;DR

Nano-drone clutter navigation must fit tight sensing and compute budgets, and fully onboard event-based goal-directed navigation on a commercial microcontroller remained open. The paper develops an analytic event proxy that generates events from scene geometry and ego-motion without rendering. Its policy transfers zero-shot to simulation and real flight, including an 82% average success rate across three real arenas.

Abstract

from arXiv · show

Event cameras offer low latency, high dynamic range, and milliwatt-level power consumption, making them ideal for agile robots such as tiny drones navigating through cluttered environments. However, existing simulators generally render photorealistic images before converting them into events, making large-scale reinforcement learning prohibitively slow. We propose an analytic event simulator that exploits what event cameras fundamentally measure - moving contrast edges - to synthesize event streams directly from scene geometry and ego-motion without visual rendering. Our simulator generates a 64x64 event observation in 17 microseconds on a single GPU - over 400x faster than a state-of-the-art rendering pipeline - enabling end-to-end reinforcement learning for goal-directed obstacle avoidance to converge in under 30 minutes - three orders of magnitude faster than training on rendered events. Policies transfer zero-shot to physically based event simulations (ray-traced Isaac Sim). Our policy reaches 83% success in cluttered courses unseen during training (mid-difficulty, traversability probability approx. 0.57), at 2.4 m/s average goal velocity, and retains 70% in a simulated texture-dense indoor hall. Furthermore, we quantized the policy to a 1.2 MB int8 network running fully onboard a 56-gram Crazyflie equipped with a Prophesee GenX320 event camera and an OpenMV N6 microcontroller board. Without any fine-tuning, external processing, or positioning, the policy achieves an average success rate of 82% (5/5, 5/6, 4/6 flights) across three real-world obstacle courses of increasing difficulty. Code: https://github.com/tudelft/rendering-free-event-nav

I. INTRODUCTION

Nano-drone navigation is constrained by sensing and compute budgets, while rendered event simulation limits reinforcement-learning throughput. The paper introduces an analytic event proxy and reports zero-shot transfer to simulation and onboard flight.

I. INTRODUCTION

  • Sub-100-gram drones cannot readily carry depth cameras or embedded GPUs, while coarse 8×8-zone ToF arrays may miss thin wires and slender poles.The sensing, processing, and carrier payload budget is only a few dozen grams and roughly a watt.Source: §I · p. 1
  • Conventional event simulators render image sequences before converting them to events, throttling the millions of closed-loop transitions required for end-to-end RL.Event cameras offer microsecond temporal resolution, high dynamic range, and milliwatt-level power consumption.Source: §I · pp. 1–2
  • 17 µs per frame and under 30 minutes of RL training demonstrate a rendering-free proxy that trains avoidance policies three orders of magnitude faster than rendered-event pipelines.The proxy generates events from scene geometry and ego-motion without visual rendering.Source: §I · p. 2
  • 83% success in unseen cluttered simulated rooms and 100%, 83%, and 67% success across three real environments demonstrate zero-shot transfer to simulation and flight.The real-flight policy runs onboard a 56-gram Crazyflie as a 1.2 MB int8 network.Source: §I · p. 2

II. RELATED WORK

Prior event-based flight work spans indirect pipelines that estimate intermediate quantities and direct event-to-command policies. Fully onboard, goal-directed clutter navigation on a commercial microcontroller remained an open capability.

A. Event-based drone flight

  • Prior event-based flight methods include indirect flow-, time-to-contact-, or depth-based pipelines and direct event-to-command policies.Examples include event optic-flow focus-of-expansion estimation, self-supervised monocular event depth, and direct policies trained or fine-tuned with event data.Source: A. Event-based drone flight · p. 2
  • Existing event-based systems often run off-board or on computers drawing 10 W or more, beyond nano-drone power and payload budgets.A neuromorphic Loihi system demonstrated onboard event-based optical-flow flight, landing, and hovering, but not goal-directed navigation through clutter on a commercial microcontroller.Source: A. Event-based drone flight · p. 2
  • Fully onboard, goal-directed navigation through clutter on a commercially available microcontroller remained an open capability.The paper presents this as the capability it demonstrates.Source: A. Event-based drone flight · p. 2

B. Event camera simulation

Existing event simulators generally render images before generating events, while render-free alternatives have not provided closed-loop 3D event generation for end-to-end RL. The paper connects analytic event simulation with microcontroller-scale learned flight.

B. Event camera simulation

  • Existing event simulators either replay fixed rendered footage or couple an event model to a renderer, so they must produce images before events.EVIS accelerates rendering in Isaac Sim and serves as the paper’s physically based reference simulation.Source: B. Event camera simulation · p. 2
  • The analytic proxy generates closed-loop events without images or an offline rendered dataset, then validates transfer to physical event simulation and a real sensor without fine-tuning.This contrasts with prior work that avoids event rendering during policy search by learning from a one-time rendered dataset and privileged state.Source: B. Event camera simulation · p. 2

C. Autonomous nano-drones and learned flight

  • The proxy pipeline represents scenes as 3D edge points on depth discontinuities, projects visible points under ego-motion, applies event-mass and polarity rules, and accumulates a 64×64 DIFF3D frame.The saturating trigger’s knee is fitted to real sensor recordings; polarity combines a hashed contrast class with the edge’s covering or revealing side.Source: Figure 2 · p. 3
  • Prior nano-drone systems use frame or range sensing and supervised perception or hand-designed control; none combines end-to-end RL, event vision, and fully onboard microcontroller inference at nano-drone scale.Earlier learned vision-based flight commonly assumes rendered or real visual data and desktop- or Jetson-class compute.Source: C. Autonomous nano-drones and learned… · p. 3

III. METHODOLOGY

The methodology replaces rendered images with an analytic event proxy built from moving scene edges and ego-motion. It models the sensor’s block-level event observations, polarity, and secondary effects, then calibrates the proxy against real sensor recordings.

III. METHODOLOGY

  • The pipeline trains reinforcement-learning policies in an analytic event proxy, validates them in physically based event simulation, and deploys quantized policies on a microcontroller.The three stages are summarized in Fig. 1.Source: §III · p. 3

A. Event proxy generation

  • The proxy represents scenes as 3D edge points on depth discontinuities and projects their motion into image-plane streaks from camera velocity and rotation.The streak is exact under translation and uses a small-angle approximation under rotation; at the stated maximum yaw rate, its arc–chord deviation is below 10^-2 px over 20 ms.Source: A. Event proxy generation · pp. 3–4
  • Each 64×64 output cell aggregates signed event mass, with a calibrated saturating ramp modeling native-pixel threshold crossings and block-mean downsampling.Gain g represents the count deposited by a fully firing edge point, while κ sets the displacement at which the native block is fully swept.Source: A. Event proxy generation · p. 4
  • Polarity combines a spatially hashed contrast class with edge-covering or revealing direction, producing coherent signs as edges move.A contrast class is fixed in world space, while outward normals determine whether an edge covers background or reveals it.Source: A. Event proxy generation · p. 4
  • Foreshortening, occlusion, analytically shaded textured ground, sketches of clutter props, and per-episode randomization model additional frame effects.Domain randomization varies gain, firing knee, polarity grain, ground-texture fill, and dilation.Source: A. Event proxy generation · p. 4
  • Calibration fits five proxy parameters to five event-frame statistics measured on real GenX320 recordings under matched motion.The check uses a single 10-second hand-moved office recording, not flight or task data, and is repeated after proxy changes.Source: A. Event proxy generation · p. 4

B. Reinforcement learning

The reinforcement-learning task trains a policy to navigate from randomized starts to a far-side goal in clutter, using event frames and goal and motion state. Training uses the deployed decision rate and reaches high success quickly on the vectorized proxy.

B. Reinforcement learning

  • The task sends the policy from a randomized spawn in a 22×22 m room through randomized clutter to a goal on the far side.Episodes end on reaching the goal within 1 m, collision, leaving the room, or timeout.Source: B. Reinforcement learning · p. 4
  • The policy receives three stacked 64×64 DIFF3D frames plus goal bearing, yaw-rate, and speed state, and outputs yaw rate and forward speed.Yaw rate is limited to 1 rad/s, and forward speed is bounded by the configured maximum.Source: B. Reinforcement learning · p. 4
  • Training uses 13.3 Hz decisions because the deployed inference and frame-assembly latency sets the stride between successive stacked observations.The rate is realized in simulation with alternating 7/8-tick holds at 100 Hz.Source: B. Reinforcement learning · p. 4
  • The reward combines goal-distance progress and potential-based clearance and alignment shaping with penalties for crashes and a reward for reaching the goal.Heading pressure is gated to open space, away from obstacle avoidance.Source: B. Reinforcement learning · p. 4
  • 25/30 shadow-scene episodes and 21/30 texture-scene episodes reached the goal in the simulation evaluation.The texture scene adds background events absent from the shadow scene, while obstacle layouts are identical across appearance conditions.Source: Figure 3 · p. 5
  • 80% training success was crossed within 8 minutes, and the 40M-step run completed in 34 minutes on one RTX 4090.The vectorized proxy sustained 19–21k environment steps/s.Source: B. Reinforcement learning · pp. 4–5

C. Sim-to-sim validation

Sim-to-sim validation compares the proxy with ray-traced Isaac Sim scenes and real GenX320 observations using a shared frame format. The proxy closely matches per-pixel shape statistics, while event density remains below reality because ambient texture is omitted.

C. Sim-to-sim validation

  • Isaac evaluation uses ray-traced HDR keyframes, motion-vector warping, and a noisy log-intensity threshold model before converting events to the shared firmware-style observation format.A single scale is fitted so rendered active-pixel p90 magnitude matches a real GenX320 recording.Source: C. Sim-to-sim validation · p. 5
  • Native-resolution event generation preserves thin obstacles because threshold-nonlinear firing can make a one-pixel-wide pole disappear when rendered directly at 64×64.Events are generated at 320×320 before block-mean downsampling.Source: C. Sim-to-sim validation · p. 5
  • The Isaac scenes are held out from proxy training, including tree fields, textured walls, and solid-geometry props.The evaluation therefore tests transfer to rendered scenes not used in proxy training.Source: C. Sim-to-sim validation · p. 5
  • The proxy matches polarity balance, magnitudes, and adjacency closely, but its event density is lower than reality because it omits ambient texture.The texture Isaac condition tracks the real recording more closely, while even the shadow renderer remains below half of real density.Source: C. Sim-to-sim validation · p. 5, Table II · p. 5

D. Real-world deployment

The policy runs onboard a 56 g Crazyflie through an N6 companion board, with control split between the policy and the drone’s stabilization firmware. The deployed network takes event observations and motion-goal context, then sends normalized yaw-rate and speed commands.

D. Real-world deployment

  • The deployed policy is a 1.2 MB int8 binary, compiled for the STM32N6 Neural-ART NPU without fine-tuning or architecture changes, and its actions match the reference in replay checks.Convolutions and dense layers run on the NPU; frame assembly, normalization, and output tanh run on the Cortex-M55.Source: D. Real-world deployment · p. 5
  • The N6 runs the policy while the Crazyflie’s STM32F405 handles low-level stabilization; a two-wire UART link carries messages and discards packets with invalid checksums.The Crazyflie maps normalized policy outputs to yaw-rate and forward-speed setpoints, with physical scaling and stabilization handled onboard.Source: D. Real-world deployment · pp. 5–6
  • The assembled flight platform totals 56 g and includes the Crazyflie base, battery, N6 board, GenX320 event camera, camera mount, flow deck, cables, and boost converter.The flow deck supports velocity estimation.Source: Figure 4 · p. 6
  • The policy receives event images plus a four-value context vector for goal bearing, yaw rate, and speed, with no map, waypoints, obstacle state, or external positioning.The N6 runs at 13.3 Hz and consumes the latest context; the Crazyflie holds the last action while its inner loop runs at 500 Hz.Source: D. Real-world deployment · p. 6

A. Evaluation protocol and metrics

The evaluation compares methods on identical procedurally generated layouts, seeds, dynamics, and action interfaces. It tests both sparse shadow scenes and texture-dense scenes, using AvoidBench metrics.

A. Evaluation protocol and metrics

  • Four methods are evaluated on the same 30 episodes, with identical layouts, spawns, goals, seeds, dynamics, and a shared action interface.The episodes comprise three fixed-seed batches of ten, each with a distinct procedurally generated layout and props; traversability is approximately 0.57.Source: A. Evaluation protocol and metrics · p. 6
  • Each layout is tested in shadow scenes with bare surfaces and cast-shadow events, and texture scenes with court markings, wall panels, and window bands under diffuse light.The paired appearances allow comparison under different visual event conditions while holding layouts fixed.Source: A. Evaluation protocol and metrics · p. 6
  • The protocol reports success rate, average goal velocity, path optimality, mission progress, and environment traversability using AvoidBench metric definitions.Path optimality is excess travel distance over the A*-shortest free path; traversability is normalized mean free-flight distance from random poses.Source: A. Evaluation protocol and metrics · p. 6

B. Zero-shot transfer

The policy transfers zero-shot to unseen Isaac scenes and retains performance in a texture-dense appearance condition, while comparisons include depth, reactive, and rendered-event baselines. Real-arena success rates are also reported for the unchanged checkpoint.

C. Baseline comparisons

  • The same checkpoint records 5/5, 5/6, and 4/6 real-arena successes, with all flights fully onboard.The figure identifies the courses as poles and boxes, an extended course with block towers, plants, and screens, and a compact arena with tight passages.Source: B. Zero-shot transfer · p. 6, Figure 5 · p. 7

B. Zero-shot transfer

  • Under the texture shift, the depth policy changes from 100% to 97% success while the event policy changes from 83% to 70%, retaining its margin over event-based alternatives.The shift doubles active pixels from shadow to texture; the int8 policy records 19/30 versus 21/30 for fp32, which the authors say is not a meaningful contributor to the texture gap.Source: B. Zero-shot transfer · pp. 6–7
  • At traversability Pτ ≈0.57, the policy reaches 83% success and 2.38 m/s average goal velocity in shadow scenes, and 70% success in texture scenes.These are zero-shot results for the same checkpoint; Table IV summarizes the shared Isaac protocol and its comparisons.Source: B. Zero-shot transfer · p. 6, Table IV · p. 7

C. Baseline comparisons

  • The privileged-depth policy reaches 100%/97% success, while the event policy reaches 83%/70%; the hand-designed event controller reaches 13%/17% in shadow/texture scenes.The depth policy uses noiseless privileged geometry; reactive control survives only the sparsest layouts, while the learned policy handles dense clutter.Source: C. Baseline comparisons · p. 7
  • Rendered-event training reaches 40%/53% success after a 24-hour GPU-day, versus the analytic-proxy policy’s 83%/70% after 34 minutes; the rendered pipeline requires over 400× more time per event frame.The rendered learner collects 195k transitions versus 40M for proxy training; its sample efficiency is high, but renderer throughput limits its training scale.Source: C. Baseline comparisons · p. 7

D. Real-world flight results

The unchanged int8 policy completes flights in three real obstacle arenas with all computation onboard. The conclusion reports a zero-shot progression from unseen Isaac scenes to hardware and identifies faster control, full attitude, and dynamic scenes as next steps.

D. Real-world flight results

  • The onboard 1.2 MB int8 policy reaches the goal in 5/5, 5/6, and 4/6 flights across three arenas, with computation fully onboard.Flights require crossing the course without touching an obstacle; the arenas include poles, boxes, cylinders, plants, and banners.Source: D. Real-world flight results · p. 7
  • In the densest arena, both crashes occur in cluttered pockets, consistent with the failure pattern in dense-room simulation episodes.Organic and structural props reduce success at a rate consistent with the Isaac props-condition results.Source: D. Real-world flight results · p. 8

V. CONCLUSION

  • The conclusion reports 83% success in unseen ray-traced event simulation, 70% under a visual-realism shift, and 82% average success across three real arenas for the 56 g Crazyflie.It identifies full attitude, faster control rates, and dynamic scenes as natural next steps.Source: §V · p. 8
Loading 2610.37543v1…