PaperlayerView exact revision on arXiv

Source-linked AI summary

SCET-CBF: Speed-Calibrated Soft Guidance for One-Step Dynamic Navigation

Xiaoyu Han, Dongkun Han

arXiv:2610.33471v1cs.RO

TL;DR

For steering-limited navigation, braking can reduce closing speed while also slowing the lateral response needed to evade later obstacles. SCET-CBF uses an escape-based speed allowance for soft guidance within segment and braking constraints; paired studies link improved dense passage to speed feedback and early activation.

  • Problem

    For a steering-limited bicycle, braking can lower closing speed while reducing yaw and lateral response, and rebuilding speed takes time under bounded acceleration.

    Source: §I · p. 1

  • Method

    SCET-CBF uses a short bicycle escape model to derive an approach-speed allowance whose speed derivative guides a one-step quadratic program within segment and braking constraints.

    Source: §I · p. 1

  • Results

    Paired studies link improved dense passage to speed feedback and early activation.

    Source: §VI · p. 8

  • Takeaways & Limitations

    Segment checks protect accepted motion, while lateral limits and bounded errors define the evaluated operating range.

    Source: §VI · p. 8

    The common single-step inflation at prediction nodes provides no full-horizon robust tube, and the primary runs exclude computation delay.

    Source: F. Limitations and failure analysis · p. 8

Abstract

from arXiv · show

Control barrier functions (CBFs) support collision avoidance, yet dense dynamic encounters can leave a vehicle with no feasible command. Geometric refinements and soft constraints help preserve motion choices. Using these choices also requires accounting for the vehicle's changing turning response. For a steering-limited bicycle, braking can lower closing speed while slowing the lateral motion needed before an obstacle arrives. We propose SCET-CBF, a speed-calibrated soft guidance law for one-step dynamic navigation. A short lateral escape maneuver gives an allowable approach speed. Its speed derivative allows a quadratic program to weigh both effects of braking when choosing acceleration and steering. Soft penalties accommodate competing obstacle preferences within the input set permitted by segment and braking constraints. We derive the guidance and establish protection over each accepted motion segment, together with conditional braking invariance for a single obstacle. On 2500 paired tasks with a 60 s deadline, SCET achieves 72.8% completion, compared with 49.8% for CBF/velocity-obstacle control and 69.8% for an enhanced 1 s CBF-MPC. At 100 obstacles, SCET and the enhanced MPC complete 52.4% and 33.6% of tasks. Paired studies identify the contributions of speed feedback and early activation. Further evaluations examine the operating range under lateral acceleration limits and bounded motion and execution errors.

I. INTRODUCTION

Dense moving obstacles can exhaust bounded commands, and braking can reduce closing speed while also slowing the lateral response needed for later evasions. SCET-CBF uses a bicycle escape-based speed allowance to guide choices within segment and braking constraints.

I. INTRODUCTION

  • Dense encounters can exhaust available commands, while clearing one obstacle may leave little room for the next.Velocity obstacles identify dangerous relative motion, and control barrier functions express safety conditions as input constraints.Source: §I · p. 1
  • Braking lowers closing speed but reduces yaw rate and lateral response, which scale with v and v^2.Under bounded acceleration, rebuilding speed takes time, so slowing for one obstacle can leave too little response for the next.Source: §I · p. 1
  • SCET-CBF derives an approach allowance from a short bicycle escape model and retains its speed derivative to weigh both effects of braking.Soft obstacle guidance operates within segment and braking constraints, allowing competing preferences to yield while preserving hard-feasible commands.Source: §I · p. 1

A. Vehicle and execution model

  • The sampled bicycle model uses a directly commanded sideslip surrogate, with steering determining current displacement and acceleration changing the next speed.Updates use 0.05 s Euler steps and linear motion between successive positions.Source: A. Vehicle and execution model · p. 2

B. Design origins and closest methods

SCET-CBF builds on soft velocity-obstacle guidance and barrier-based braking, deriving its maneuver allowance from a bicycle escape maneuver. Segment checking recovers the full safe steering domain under the sampled model before speed and braking constraints further restrict the commands.

B. Design origins and closest methods

  • MAVO uses velocity-obstacle geometry as a soft objective with a hard braking barrier, while DPCBF adapts a hard parabolic boundary with distance and relative speed.SCET extends these designs with a soft approach allowance derived from a bicycle maneuver.Source: B. Design origins and closest methods · p. 2
  • SCET derives its allowance and activation time from a short bicycle escape maneuver, unlike SC3BF’s shifted hard collision-cone boundary for a continuous-time single-obstacle bicycle model.SCET uses the allowance’s speed sensitivity as soft guidance among multiple obstacles, with segment checking and braking constraints as protection.Source: B. Design origins and closest methods · p. 2

III. THE SCET-CBF UPDATE

  • Segment geometry and braking define the available input set before soft guidance ranks choices using the vehicle’s turning response.Figure 2 relates physical wheel angle to the steering surrogate used by the guidance model.Source: §III · p. 2, Figure 2 · p. 2

A. Step 1: steering commands that clear the next segment

  • The method recovers the entire continuous safe-steering domain by finding branch or safety-change roots and testing intervening intervals and boundaries.The relevant roots are at most quadratic in steering.Source: A. Step 1: steering commands that… · p. 2
  • With bounded constant steering error, the method erodes each safe interval and inflates obstacle radii by a bound on relative path error over the sample.The stated coverage assumption observes obstacles within 15 m, and the actual paths must obey the error bounds used to construct the inflation.Source: A. Step 1: steering commands that… · pp. 2–3

B. Step 2: speed limits and braking margin

A braking margin compares obstacle separation with the distance needed to reduce radial approach, jointly constraining acceleration and steering. The resulting hard input set intersects safe steering intervals with acceleration and next-speed limits.

B. Step 2: speed limits and braking margin

  • The braking margin compares separation with the distance needed to reduce radial approach, and its condition jointly constrains acceleration and steering.Theorem 2 derives a conservative lower bound on combined radial control capability; relative velocity covers constant-velocity moving obstacles, including static ones.Source: B. Step 2: speed limits and braking… · p. 3
  • When a braking margin is negative, its row requests recovery; the enforced rows include nearly tied worst margins among processed obstacles.Processing includes the nearest ten obstacles within 15 m and those meeting the stated distance criterion.Source: B. Step 2: speed limits and braking… · p. 3
  • The hard input set intersects the steering intervals with acceleration bounds and limits that keep next speed within the allowed range.The longitudinal braking limit remains 5 m/s^2.Source: B. Step 2: speed limits and braking… · p. 3

C. Step 3: selecting for continued maneuverability

SCET-CBF derives an approach allowance from a short lateral escape maneuver and uses its speed sensitivity to guide acceleration and steering. The guidance activates before a declining margin is exhausted, accounting for the time needed for turning response to develop.

C. Step 3: selecting for continued maneuverability

  • The steering-limited bicycle’s quarter-turn displacement sets a proxy scale for the lateral escape allowance, which is matched to constant lateral acceleration over the turn time.The maneuver model uses maximum steering to derive yaw rate and quarter-turn time, then uses the heading-rotation contribution to set the proxy scale.Source: C. Step 3: selecting for continued… · p. 3
  • For nonfrontal encounters, the margin uses relative approach, sideways speed, and total relative speed; existing sideways motion is credited, and positive margin means approach is within the allowance.The tangency construction has a unique minimum; the generalized allowance applies when distance exceeds 1.02 times the avoidance radius.Source: C. Step 3: selecting for continued… · p. 3
  • Braking changes both approach speed and turning allowance, so the QP uses the allowance’s speed derivative when evaluating the preference.For a stationary frontal obstacle, the preference can favor acceleration at d = 2 m but favor slowing at d = 0.8 m.Source: C. Step 3: selecting for continued… · p. 3
  • The soft row activates before the current approach preference is lost, with earlier warning when turning is slower.Its advance combines a one-update margin estimate with physical maneuver time; approaching obstacles also receive time-to-collision weights.Source: C. Step 3: selecting for continued… · p. 3
  • The evaluation retains fixed guidance parameters, while the implementation floors speed for the allowance and switches definitions without smoothing near the avoidance radius.The 60 s study uses the same parameters; the implementation uses max(v, v_min) and an unsmoothed switch at 1.02r_cj.Source: C. Step 3: selecting for continued… · p. 4

D. Soft guidance within the hard input set

Independent soft preferences let the QP accommodate conflicting obstacle guidance without removing inputs from the hard feasible set. It selects a lowest-cost validated command across the feasible input components.

D. Soft guidance within the hard input set

  • Independent slacks let the QP balance conflicting obstacle preferences while retaining every input in the hard set.For any hard input, choosing each slack to cover its row violation satisfies the soft rows; slack cost measures opposition to the corresponding preference.Source: D. Soft guidance within the hard input… · p. 4
  • The update forms segment-safe steering intervals, intersects them with speed and braking constraints, then solves the soft problem on each nonempty component.If a solve fails, it projects the nominal input and validates it; returned candidates undergo bounds, hard-row, and segment checks.Source: D. Soft guidance within the hard input… · p. 4
  • A violated preference pulls the input along its margin sensitivity, with violation and urgency setting pull strength; speed feedback changes the acceleration component.A satisfied preference adds no correction.Source: D. Soft guidance within the hard input… · p. 4
  • Each feasible component is a convex polygon, and comparing optimized costs across nonempty components selects the best input over their union.The equivalent objective after eliminating slacks makes the preference mechanism explicit.Source: D. Soft guidance within the hard input… · p. 4

A. Protection of the current segment

The paper establishes clearance over each accepted sample segment under stated error and obstacle-coverage assumptions, and proves conditional braking-margin invariance for one obstacle under bounded conditions.

A. Protection of the current segment

  • Under Assumption 1, an accepted command stays outside physical obstacle discs throughout the sample when its final check includes error bounds and a positive physical cushion.The segment check covers the whole relative path; steering erosion, path-error inflation, and the reach bound for omitted obstacles support the guarantee.Source: A. Protection of the current segment · p. 4
  • With exact arithmetic, zero path and steering errors, and nonempty A(v), the returned interval union is empty exactly when no held input clears the declared radius for the sample.Root partitioning covers sign and branch changes, so no admissible steering is omitted.Source: A. Protection of the current segment · p. 4

B. Conditional braking margin invariance

  • For one constant-velocity obstacle, Eq. (5) preserves the braking-margin set when its stated radius, speed, and braking-parameter conditions hold.The theorem requires r_cj ≥ 0.4 m, v ≥ 0.2 m/s, and 0 < a_B < 0.0600 m/s²; the comparison argument establishes invariance while domain and input conditions hold.Source: B. Conditional braking margin invariance · pp. 4–5
  • The conservative braking bound is tied to the bicycle’s input limits, including least-favorable low-speed orientations; multiple obstacle rows and next-speed limits are evaluated jointly later.The proof’s orientation split compares acceleration and steering contributions; the stated example gives lower bounds 0.0600 and 0.06056 m/s² against a_B = 0.055 m/s².Source: B. Conditional braking margin invariance · p. 5

C. Optional continuation after task termination

An optional committed-plan mode preserves a verified continuation after rejection or task termination, subject to verifier bounds and plan-availability conditions. The evaluation protocol compares controllers on shared seeded dynamic-obstacle scenarios with fixed configurations and reports timing and uncertainty procedures.

C. Optional continuation after task termination

  • The optional continuation mode accepts an input only after committing its verified execution and a valid successor plan; on rejection or task termination, it executes the saved plan.Validity depends on verifier bounds, available unexecuted suffixes, and replacements matching time and execution-state anchors.Source: C. Optional continuation after task… · p. 5

V. EVALUATION

  • The evaluation tests sustained passage, guidance mechanisms, and operating range across shared-seed scenarios with obstacle counts from 10 to 100 and a 60 s goal deadline.Each controller faces 500 seeds at each obstacle count; obstacles subsequently disperse on an unbounded plane.Source: §V · p. 5, A. Protocol and direct comparisons · p. 5

A. Protocol and direct comparisons

  • Direct comparisons use a shared plant, goal, input bounds, observations, and segment evaluator, with controller configurations selected or fixed as described before the 2500-world comparison.Reported procedures include 95% pointwise bootstrap intervals, Holm-corrected seed-label tests, and full-call timing on ten shared worlds per method.Source: A. Protocol and direct comparisons · p. 5

B. Sustaining passage as obstacle density increases

Across paired navigation tasks, SCET outperformed one-step CBF-QP baselines and achieved a larger advantage over enhanced MPC in 100-obstacle encounters. Ablations also show gains associated with turning calibration, speed feedback, and early activation.

B. Sustaining passage as obstacle density increases

C. Which guidance choices improve passage?

  • Guidance increased completions from 295 with braking alone and 224 with segment protection alone to 1820; braking added +5.1 pp when guidance was present.The counts compare successive additions of braking and guidance to the hard segment-protection layer.Source: C. Which guidance choices improve… · p. 6
  • Turning calibration yielded 364 completions versus 37 for the reproduced DPCBF frontal rule on 500 paired worlds, a +65.4 pp difference.The comparison replaced the frontal rule and its derived feedback and activation quantities at fixed weights.Source: C. Which guidance choices improve… · p. 6
  • Full versus frozen speed feedback completed 74/31 worlds without early activation and 130/17 with it; early activation raised full-feedback completions from 74 to 130.With early activation off, freezing feedback raised speed-floor occupancy from 1.3% to 22.1%; distance-based allowances had identical full/frozen outcomes.Source: C. Which guidance choices improve… · p. 7

D. Guidance under lateral-acceleration limits

SCET adjusts its turning allowance to lateral-acceleration limits and evaluates guidance under changing obstacle motion, actuator dynamics, and optional continuation. The results show both gains from matching the allowance to available turning response and trade-offs under bounded errors.

D. Guidance under lateral-acceleration limits

  • At 0.1g/0.25g/0.5g, matching SCET's allowance to available turning response adds 24.5/12.0/−2.5 pp over its original uncapped allowance.The first two gains retain Holm-adjusted p < 0.05.Source: D. Guidance under lateral-acceleration… · p. 7

E. Bounded motion and execution errors

  • SCET completes 120/100 tasks with zero recorded contacts under acceleration/jump motion laws, versus enhanced MPC's 69/101.The acceleration-law difference is 25.5 pp [17.5,33.0], while the jump-law difference is −0.5 pp [−8.0,6.5].Source: E. Bounded motion and execution errors · p. 7, Table III · p. 8
  • Across 400 paired profile/world executions, including SCET's path bound reduces contacts from 50 to zero, while completion changes from 179 to 167.The largest path-error/bound ratio is 0.914; the comparison uses identical profiles and bound information.Source: E. Bounded motion and execution errors · p. 7, Table IV · p. 8
  • On 200 separate worlds with ideal plant and zero computation delay, 91 of 200 starts reach the goal before fallback, and all retain a valid continuation.All starts are admitted, and 49 short turn/speed maneuvers connect to verified straight tails.Source: E. Bounded motion and execution errors · p. 8

F. Limitations and failure analysis

The analysis identifies exhaustion and numerical or modeling boundaries alongside the method's safety scope. The conclusion limits the tested range and points to physical validation and richer prediction as future work.

F. Limitations and failure analysis

  • The 670 primary SCET exhaustions comprise 513 empty steering sets and 157 hard-intersection losses.Continuing with straight braking to the speed floor gives 651 later contacts under the constant-velocity model; reaching E supplies no stopping certificate.Source: F. Limitations and failure analysis · p. 8
  • At N = 10, SCET/MAVO/MPC H10 complete 468/491/486; removing SCET's braking row raises its count to 497 but reduces N = 100 completion from 262 to 171.The braking layer therefore has a density-dependent cost and benefit.Source: F. Limitations and failure analysis · p. 8
  • The single-step inflation provides no full-horizon robust tube, primary runs exclude computation delay, and physical validation remains necessary for tire transients, friction sharing, sensing, and servo calibration.Future continuation certificates also require predictions valid over their entire execution.Source: F. Limitations and failure analysis · p. 8

VI. CONCLUSION

  • The conclusion places the tested range at segment checks, lateral limits, and bounded errors, with physical tests, measured steering dynamics, richer prediction, and safe continuation left for future work.Source: §VI · p. 8

Related papers

Loading 2610.33471v1…