acceptodds
Under review as a conference paper at ICLR 2027

Gated Learned-Dynamics Planning for Efficient UAV Obstacle Avoidance

Abstract

Learned dynamics models let an agent anticipate candidate actions, but invoking model-based planning at every control step can waste computation when reactive behavior is sufficient. We study when short-horizon online imagination is necessary for UAV obstacle avoidance in a controlled cluttered 3D corridor. A compact learned kinematics model supports receding-horizon target-grid planning; a fixed-clearance hybrid invokes this planner only near hazards and otherwise uses behavioral cloning. Across 60 seeded evaluation maps, reactive control and behavioral cloning succeed on 7% and 32% of trials, while always-on learned planning and the gated hybrid succeed on 85% and 87%, respectively. The hybrid reduces mean planning time from approximately 12.0\ms to 6.4\ms, with overlapping success intervals. On the 40 maps withheld from gate selection, success is 33/40 and 34/40. Call-level measurements attribute the time reduction to issuing about half as many of the same rollouts; a lateral boundary override inside the gate is rarely active. Development-holdout attempts to distill planned actions into a reactive policy do not recover online planning. An additional PPO training attempt is inconclusive because its selected checkpoint is the behavioral-cloning warm start, before policy-gradient updates. These results provide a controlled study of selective model-based computation, while exposing the limitation that runtime clearance and imagined-pose costs use privileged geometric information.

Primary area:
applications to robotics, autonomy, planning
Keywords:
learned dynamics, model-based planning, reinforcement learning, UAV navigation, obstacle avoidance, robotics

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_qtz4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

In the paper "Gated Learned-Dynamics Planning for Efficient UAV Obstacle Avoidance," the authors propose a hybrid controller that switches between a behavioral cloning policy and learned-dynamics model predictive control (MPC) based on a fixed obstacle-clearance threshold. They demonstrate comparable obstacle-avoidance success to always-on MPC while reducing average planning time from 12.0 ms to 6.4 ms in a simulated UAV environment.

Overall, I find that the paper poses an interesting and timely question about when computationally expensive planning is necessary, but does not meaningfully answer it. The proposed switching criterion is a manually selected distance threshold, with little analysis of when this threshold is appropriate or whether the resulting computational savings are practically meaningful. Answering the broader question would require substantially more than what the authors present.

Critical strengths

The question of when to invoke computationally expensive planning is relevant to resource-constrained autonomous systems. The paper also provides controlled comparisons and several ablations, including an analysis of the switching threshold and computational costs.

Critical weaknesses
  1. The central research question remains unanswered. The authors ask when online planning is necessary, but effectively assume that obstacle clearance determines the answer. A fixed clearance threshold tells us little about whether a reactive policy can safely handle the current situation. This depends on the vehicle's velocity, dynamics, control authority, obstacle geometry, and the time required to execute an avoidance maneuver. The authors do not establish when switching becomes necessary, or even whether the selected threshold generalizes beyond their particular simulation. This is arguably the most interesting part of the problem, yet it is largely left unaddressed.

  2. The computational motivation is unconvincing. The primary result reduces average planning time from approximately 12 ms to 6 ms, but the authors do not establish why this reduction matters for their application. It is unclear whether this produces any meaningful improvement in real-time control performance or computational feasibility.

  3. The related work does not adequately establish the contribution. Event-triggered control and event-triggered MPC have studied when expensive optimization should be invoked. The authors briefly acknowledge this literature but do not meaningfully distinguish their approach or compare against relevant triggering strategies.

  4. Several assumptions limit the conclusions. The switching criterion and planning objective rely on privileged ground-truth obstacle geometry, while the UAV simulator uses simplified kinematics rather than realistic flight dynamics. What is the justification for these and why are they chosen?

Questions for authors
  1. What determines when the system should switch from reactive control to MPC? Why is geometric clearance an appropriate criterion, independent of velocity, dynamics, or the time required to avoid an obstacle?

  2. What practical computational constraint motivates reducing mean planning time from 12 ms to 6 ms? Given that individual planner calls take approximately 47 ms, how does the proposed approach improve real-time feasibility?

  3. How does the proposed method differ substantively from existing event-triggered MPC approaches? Would comparisons against alternative triggering conditions or reduced-frequency MPC lead to similar results?

Reviewer_LeVDOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper investigates whether a drone that plans with a learned dynamics model needs to keep planning all the time, or only when it comes close to obstacles. In a simple kinematic corridor simulator, a behavioral-cloning policy flies in open space and a short-horizon planner using a learned kinematics model takes over when the distance to the nearest obstacle drops below a fixed threshold. The authors show that this switching controller is as successful as planning all the time, while calling the planner about half as often.

Contributions:

  • They present a seeded corridor benchmark tool (no code?): reactive vs behavioral cloning vs learned-model planning.
  • They propose a clearance gate: heavy planner only near obstacles -> same success, half the planner calls.
  • They show that reactive approach fails to match the planner.
Critical strengths
  • 50% saving in CPU, same results
  • careful statistics, honest: no equivalence claim
  • timing: saving = fewer planner calls, not cheaper ones
  • limitations stated openly (ground-truth geometry)
  • protocol well documented: seeds, checkpoints, commands (but code could not be found)
Critical weaknesses
  • simulation only, simplified kinematics (sphere, no rigid-body flight dynamics) -> no real flight
  • MAIN WEAKNESS: How I read it, the planner gains information from the map (eq 2. c(s_t) (=real learned imagination?)
  • gate system cuts average compute, not worst case. some calls are longer than control period
  • gate = distance, should be time to collision. vehicle speed could not be found. (our estimate: 61 m track in 8 s = approx 7 m/s, so 3.5 m = ~0.5 s before contact vs 1.4 s horizon)
  • one simplified corridor only, AvoidBench results left out
  • limited novelty for the conference: planning near obstacles only = common practice, no comparison with existing trigger rules
  • Noise ablation -> 1 trial/map too few
Questions for authors
  • \tau in a=(...., ...,\tau) definition not be found?
  • Track A referenced before defined. Track B only in Figure 1, never evaluated?
  • Proximal Policy Optimization: validation seeds 41-50 inside training seeds 1-64?
  • Appendix E: bc_mlp_40pct + bc_wm_mlp. explain the difference. 40% is success rate?
  • always-on planner + true geometry still fails ~15% of maps: failure analysis?
  • Figure 3/4 code names unclear: (wm_rollout, bc_wm). 'locked': explanation could not be found?
  • Appendix C: 99 m gate? I thought it was only 65m?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.