PaperlayerView exact revision on arXiv

Source-linked AI summary

ClearNav: Perception-Aligned Quadrotor Flight Using Differential-Physics

Aditya Azad, Arthur Huang, Guanrui Li

arXiv:2610.17419v1cs.ROcs.AIcs.LG

TL;DR

Quadrotors in confined clutter must remain safe despite limited onboard sensing and compute, while existing differentiable-physics policies struggle with deliberate avoidance. ClearNav adds reward terms for obstacle proximity, heading, and lateral motion. Across 18 held-out static scenarios, it raises success from 90.3% to 97.1% and cuts collisions by 65%; the reported effect on moving obstacles is not significant at suite level.

  • Problem

    Differentiable-physics policies struggle with deliberate multi-step avoidance in clutter, while confined flight also constrains onboard sensing and compute.

    Source: §I · p. 1

  • Method

    ClearNav adds reward terms penalizing high speed near obstacles, heading away from a forward-looking geodesic direction, and lateral motion.

    Source: §I · p. 2, 6) Body-frame lateral-velocity penalty · p. 4, §IV · p. 7

  • Results

    On 18 held-out static scenarios, ClearNav raises mean success from 90.3% to 97.1% and cuts collisions by 65%.

    Source: §IV · p. 7

  • Takeaways & Limitations

    The policy also navigated a cluttered scene safely on a physical quadrotor, while the baseline could not.

    Source: §IV · p. 8

    The proximity and collision losses use exact omnidirectional clearance queries, while the policy observes only a single 89° depth camera.

    Source: §IV · p. 8

Abstract

from arXiv · show

Autonomous quadrotor flight in cluttered environment is an key capability for tasks like inspection and search. Navigation policies trained in simulation and transferred to hardware can fly at high speed in open space, but they can still collide as obstacle clearance shrinks. We attribute most of these collisions to a single mechanism, namely motion in directions the onboard camera does not observe. Heading that is not aligned with the direction of travel and lateral excursion during obstacle avoidance both carry the vehicle toward obstacles outside its field of view. We introduce ClearNav, a navigation policy trained with two reward terms that constrain motion to the observed direction, a forward-looking heading target and a penalty on lateral velocity. A third term penalizes speed near obstacles so that an obstacle entering the field of view remains avoidable. In simulation, across 18 held-out scenarios with 2304 paired flights per policy unseen in training, ClearNav improves success by 6.8 points, cuts the collision rate by 65% relative to the state of the arts, and flies at maximum 5m/s. For the proximity term alone under matched commanded speeds, the collision reduction still holds at 70%. The effect is significant on static scenes and consistent in direction on moving obstacles, though not significant there at the suite level, showing that safer flight in clutter is possible even on the limited sensors and compute of small drones. We also validate our navigation policy in a real-world cluttered environment, showing that it enables the quadrotor to navigate through the cluttered environment successfully and safely.

SUPPLEMENTARY MATERIAL

Flight in confined clutter must remain safe despite limited onboard sensing and compute. ClearNav adds observable reward terms for obstacle proximity, heading alignment, and strafing to a differentiable-physics policy.

I. INTRODUCTION

  • Confined geometry limits drone size, sensing, and compute, while demanding safe flight at small obstacle clearances.The introduction frames this as a need for search, rescue, and inspection.Source: §I · p. 1
  • Differentiable-physics policies struggle with deliberate multi-step avoidance, with misaligned yaw and overly wide trajectories identified as recurring failure causes.Long unrolled rollouts can be ill-conditioned, and point-mass policies may leave tangential motion unpenalized.Source: §I · p. 1
  • ClearNav adds three reward terms to a tuned differentiable-rollout objective without sacrificing degrees of freedom.The introduction identifies speed near obstacles, misaligned yaw, and excessive strafing as the penalized behaviors.Source: §I · p. 2
  • Each penalized behavior is observable at inference, so the deployed policy can act on the objective’s signals.Source: §I · p. 2
  • The contribution list includes navigation around large obstacles across varied sizes, shapes, and counts in simulation and a real-world experiment.Source: §I · p. 2
  • The reported simulation comparisons cover static out-of-distribution scenes, matched-velocity control, moving obstacles, and a cross-framework setting.Source: §I · p. 2

II. METHODOLOGY

The policy navigates from a sampled start pose to a goal through clutter using a forward-facing depth camera and proprioception. Its observations inform a recurrent policy whose parameters are optimized by backpropagating through differentiable simulator rollouts.

A. Problem formulation

  • The navigation problem uses a randomly sampled start pose and goal in an unknown cluttered environment, with only forward-facing depth and proprioception.Source: A. Problem formulation · p. 2
  • The observation includes proprioceptive state and a goal encoding that provides a capped proportional velocity command toward the goal.The state contains attitude relative to the start frame and velocity expressed in that frame, both perturbed by stepwise Gaussian noise.Source: A. Problem formulation · p. 2
  • The goal command carries no obstacle information, so avoidance must come from the depth image.Source: A. Problem formulation · p. 2
  • The action combines mass-normalized collective thrust with an absolute yaw setpoint, and the policy maps observations and recurrent state to that action.Source: A. Problem formulation · p. 2
  • Policy parameters are optimized by backpropagation through unrolled simulator dynamics, with the gradient averaged over initial conditions, scenes, and injected noise.Source: A. Problem formulation · p. 2

B. Differentiable dynamics

The dynamics model treats the quadrotor as a point mass and reconstructs attitude from thrust and yaw commands. To manage long-horizon gradient ill-conditioning, it shrinks gradients propagated backward through the rollout.

B. Differentiable dynamics

  • The model uses point-mass translational dynamics with attitude reconstructed from commanded thrust and yaw, an abstraction that can struggle when substantial yaw motion is required.Source: B. Differentiable dynamics · p. 2
  • Body attitude is reconstructed in closed form from the filtered thrust and yaw setpoint, aligning the body’s thrust axis with thrust and its forward axis as closely as possible with the commanded heading.Source: B. Differentiable dynamics · p. 3
  • Body rate is recovered by finite-differencing consecutive rotations and is directly penalized.Source: B. Differentiable dynamics · p. 3
  • Gradients propagated backward through the rollout are shrunk exponentially over a horizon to bound total gain and address long-rollout ill-conditioning.Source: B. Differentiable dynamics · p. 3

C. Environment and geodesic guidance

The training setup uses Habitat-sim scenes, exact nearest-obstacle queries, and a privileged geodesic field that shapes the objective without entering the policy. The policy combines depth, state, and goal features, and training minimizes a weighted per-step objective.

C. Environment and geodesic guidance

  • The geodesic guidance field shapes the objective but is never an input to the policy.The field is derived from a scalar time-of-arrival potential; episodes end on collision or leaving the stage, and succeed when the vehicle reaches the goal threshold.Source: C. Environment and geodesic guidance · p. 3

D. Policy architecture πθ

  • The policy uses a three-branch encoder for depth, state, and goal features, followed by a recurrent cell and linear action head.The depth branch max-pools inverse depth to preserve the closest obstacle in each cell; the branches are summed before the GRU.Source: D. Policy architecture πθ · p. 3

E. Objective

  • The per-step objective is a weighted sum of nine terms, including moving-average and exponentially weighted velocity references, body-frame velocity, and direction to the nearest obstacle.The supplied passage defines these quantities but does not specify all nine terms.Source: E. Objective · p. 3

1) Progress:

The objective rewards tracking the routed direction at commanded speed while adding penalties for collision-related behavior. Its collision safety term combines a closing-speed penalty with a soft interior barrier.

1) Progress:

  • The progress term tracks the routed direction at commanded speed, using a Huber loss on tracking error and a one-sided quadratic for exceeding the command.Windowed average velocity permits lateral excursions around obstacles without immediate penalty.Source: 1) Progress · pp. 3–4
  • The loss modifications include a velocity-aware proximity penalty, a yaw lookahead reward, and a body-frame lateral-velocity penalty.The proximity penalty discourages high speed near obstacles regardless of direction; the other terms encourage heading alignment and discourage sideways motion.Source: Figure 4 · p. 4

2) Collision:

  • The safety term combines a closing-speed penalty with a soft interior barrier that grows linearly after surface penetration.The closing-speed factor penalizes radial approach within a shell of thickness df; the barrier is negligible at positive clearance.Source: 2) Collision · p. 4

3) Smoothness:

Smoothness costs penalize acceleration, jerk, and body rate to make trajectories flyable. Heading guidance uses a lookahead reference that can follow the routed direction and turns the camera toward the path before bends.

3) Smoothness:

  • Acceleration, jerk, and body-rate penalties make trajectories flyable.The acceleration is translational and jerk is the derivative of acceleration.Source: 3) Smoothness · p. 4

4) Heading:

  • The heading reference interpolates from the goal direction toward smoothed velocity, then can use a routed direction sampled ahead to turn the heading before a bend.The goal-direction reference is gated off within 1 m of the goal, and the slerp reference is retained when the lookahead point lies inside geometry; the guidance gradient shapes training but is not a policy input.Source: 4) Heading · p. 4

5) Velocity-aware proximity cost:

The velocity-aware proximity cost penalizes speed near obstacles, while a separate body-frame lateral-velocity cost discourages fast sideways motion. The two terms address different ways motion can remain unsafe despite heading alignment.

5) Velocity-aware proximity cost:

  • The direction-agnostic proximity cost penalizes speed near obstacles, with an exponential falloff that keeps it negligible in open space.Its velocity gradient opposes current motion, with magnitude set by proximity, so the vehicle must slow down near obstacles.Source: 5) Velocity-aware proximity cost · p. 4

6) Body-frame lateral-velocity penalty:

  • The body-frame lateral-velocity penalty makes fast sideways motion costly while leaving forward and vertical velocity free.Its gradient acts only through lateral velocity, leaving small transient lateral motion in coordinated turns nearly free; unlike the proximity cost, it has no obstacle or scan reference.Source: 6) Body-frame lateral-velocity penalty · p. 4

7) Training:

Training uses a two-stage progression from goal reaching in an empty arena to cluttered scenes with time-of-arrival guidance and the full objective. The experiments use paired held-out scenes and bound claims about ClearNav’s gains to out-of-distribution settings, while noting that the comparison changes both objective terms and the training speed distribution.

7) Training:

  • Training unrolls N agents for H steps and backpropagates through dynamics and clearance in one pass, then detaches simulator and recurrent states after each update.Terminated agents are reset individually, with discount and recurrent state zeroed only for those rows.Source: 7) Training · pp. 4–5
  • Stage one teaches thrust-to-motion mapping and goal reaching in an empty walled arena; stage two carries weights forward to cluttered scenes with time-of-arrival guidance and the full objective.The training stages carry the weights and global iteration counter forward.Source: 7) Training · p. 5

III. EXPERIMENTAL RESULTS

  • The experiments ask whether ClearNav helps in-distribution, generalizes out of distribution, retains gains under speed control, and transfers to moving obstacles; the reported gains are out-of-distribution, not in-distribution.The authors report no measurable proximity-term benefit on the training distribution and a weaker, non-significant effect on moving obstacles.Source: §III · p. 5

A. Experiment Setup

  • ClearNav differs from baseline in three objective terms and its stage-one speed band, so the comparison is not a single-variable ablation; OA-only tests the proximity term alone.The full bundle is not re-run in the matched-velocity control, and the separate effects of lateral-slip and look-ahead-yaw terms are not tested.Source: A. Experiment Setup · p. 5
  • Evaluations use paired held-out scenes and seeds, define success within 0.5 m of the goal, and count geometry contact or leaving the stage as collision.The collision radius is 0.25 m, below the 0.3 m used in training; scenario-level paired success rates are tested with Wilcoxon signed-rank tests.Source: A. Experiment Setup · p. 5

B. In-distribution evaluation

On the stage-one scene family, baseline and OA-only converge to nearly identical success, collision, and speed outcomes. The proximity term neither measurably changes performance nor incurs an observed training-distribution cost.

B. In-distribution evaluation

  • ≈99% success and ≈1% collision were reached by both baseline and OA-only on the held-out stage-one scene family.The evaluation sampled 80 episodes every 100 iterations with commanded speeds of m/s.Source: B. In-distribution evaluation · p. 5
  • Both policies had identical mean speed of ≈2.5 m/s, with the proximity term neither slowing the policy nor degrading goal-reaching or training stability.The passage reports no measurable in-distribution benefit or cost for the proximity term.Source: B. In-distribution evaluation · pp. 5–6
  • Whether the full bundle matters must therefore be decided out of distribution, where its gains appear rather than on this training-distribution evaluation.Source: B. In-distribution evaluation · p. 6

C. Static generalization suite

ClearNav’s strongest reported evidence is on held-out static scenarios, where it improves success and reduces collisions despite higher mean speed. A matched-command control supports a collision reduction for the proximity term alone, while moving-obstacle results remain directionally favorable but inconclusive.

C. Static generalization suite

  • Success rose from 90.3% to 97.1% and collision rate fell from 8.2% to 2.9%, a 65% relative reduction, across 18 static scenarios and 2304 paired episodes.ClearNav was better in 17 scenarios, worse in none, and tied once; the improvement was significant with Holm-corrected p = 3 × 10^-4.Source: C. Static generalization suite · p. 6
  • The gain was largest in dense clutter, clustered layouts, and the tight arena, while the near-saturated very-sparse scenario showed no difference.The tight-cluster case was ClearNav’s weakest, with a higher collision rate plausibly dominated by its higher speed.Source: C. Static generalization suite · p. 6
  • ClearNav’s mean speed was 3.39 m/s versus B’s 2.24 m/s, a 52% increase; most of the gap came from the wider stage-one training band.The matched-command control addresses whether the safety gain can be attributed to objective terms rather than the training speed band.Source: C. Static generalization suite · p. 6

D. Controlling for commanded speed

  • With both policies commanded at U[2, 4] m/s, OA-only improved success from 90.6% to 97.2% and reduced collisions from 8.4% to 2.6%, a 70% relative reduction over 512 paired episodes.The control was partial: OA-only remained faster at 2.73 vs 2.04 m/s, and the full bundle was not re-run under matched velocity.Source: D. Controlling for commanded speed · p. 6

E. Moving obstacles

  • With obstacle separation enforced, moving-scene success rose from 64.0% to 73.9% and collisions fell from 36.0% to 26.1%, but the result was not significant at n = 11, p = 0.084.Crossing-traffic performance declined for both policies from 80–95% success at 1 m/s to 33–38% at 5 m/s, with unstable ordering; 24 episodes per scenario were too few to establish the suite effect.Source: E. Moving obstacles · p. 6

F. Cross-framework comparison

ClearNav improves held-out clutter-flight results and transfers to a physical quadrotor, while its cross-framework comparison measures whole stacks rather than isolating reward terms.

F. Cross-framework comparison

  • In the cross-framework comparison, DA achieves 61.8% mean success versus 90.3% for B and 97.1% for ClearNav, but the differing stacks prevent attributing this gap to the reward term.Table I covers 18 scenarios and identifies the best of B and ClearNav; the comparison also varies renderer, depth resolution, dynamics, termination criteria, and episode limits.Source: F. Cross-framework comparison · p. 6, Table I · p. 7

G. Real world flight

  • On a physical quadrotor running onboard, ClearNav navigates a four-obstacle scene to the goal, while the baseline grazes or crashes into the enclosure nets.The flight used a Jetson Orin NX; ClearNav first moved left around a narrow obstacle, then yawed around a wider one.Source: G. Real world flight · pp. 6–7

IV. CONCLUSION

  • On 18 held-out static scenarios, ClearNav raises success from 90.3% to 97.1%, cuts collisions by 65%, wins 17 of 18, and flies faster.Under matched-velocity control, the proximity term alone retains a 70% collision reduction; moving-obstacle effects point the same way but are not significant at suite level. The proximity term adds nothing measurable on the training distribution.Source: §IV · pp. 7–8
Loading 2610.17419v1…