acceptodds
Under review as a conference paper at ICLR 2027

Predictive Composite Guiding Vector Field for 3D Manipulator Path Following with Full-Link Obstacle Avoidance

Abstract

Safe manipulator path following requires convergence of the end effector to a desired path while maintaining clearance for all modeled links, particularly near moving obstacles and narrow passages. We propose a predictive composite guiding vector field for 3-D path following with full-link collision avoidance. An implicit-surface guiding field provides nominal motion, while capsule models and predicted capsule-aware obstacle levels account for link thickness and anticipate collisions. Moving-boundary compensation handles dynamic obstacles. A rank-discounted global index preserves dominant-threat information for event observation and rejoining, while simultaneous link corrections are reconciled by weighted damped least-squares. After threat release, event-driven safety rejoining (EDSR) modulates convergence and tangential gains under safety gating, and bounded joint commands are obtained by damped inverse kinematics. MATLAB experiments with static, moving, and narrow-frame obstacles show predictive activation and positive capsule-surface clearance; CoppeliaSim co-simulation and a UR7e experiment further illustrate the feasibility of obstacle avoidance and path recovery on simulated and physical platforms.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Guiding vector fields, manipulator path following, full-link collision avoidance, dynamic obstacle avoidance, predictive collision avoidance, event-driven safety rejoining

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_A8rhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(−1)
Soundness
1(−1)
Read review
Paper summary

This paper proposes a predictive composite guiding vector field for manipulator path following with full-link obstacle avoidance. It combines obstacle prediction, capsule-based link models, and joint-space corrections, and introduces an event-driven mechanism for returning to the path after avoidance. The method is demonstrated in simulation and on a real robot.

Critical strengths

I think the problem is useful. Keeping the end effector clear of obstacles does not mean the whole arm is safe, so considering the other links is necessary. I also like that the paper considers what happens after avoidance. The robot still needs to return to the path, and simply restoring the original controller may not handle this well. Bringing avoidance and recovery together makes sense for this task.

Critical weaknesses

What I did not fully understand is the main insight behind the method. Prediction, capsule geometry, link correction, and recovery modulation each have a clear purpose. But why combine them in this particular way? What does this combination tell us beyond the usefulness of each component? I can follow how the controller is constructed, but the paper does not make its main conceptual contribution equally clear. Different links may need conflicting motions to avoid obstacles, while returning to the path adds another requirement. The paper uses weighted corrections and gives avoidance more weight as the threat increases. But when these requirements cannot all be satisfied, what behavior should we expect? How does the design decide which requirement to compromise? Giving the weighting rules does not fully explain the reasoning behind that choice.

I also have a question about the global threat index. A lower threat value means less danger according to the index, but why does it mean that the robot should move more strongly toward the path? Threats with similar magnitudes can have different locations and directions. I understand that the individual link corrections still use this information, but how does the recovery decision based on the global index remain consistent with those corrections?

The role of event history is also not clear enough to me. What does remembering an avoidance event tell the controller that the current robot and obstacle state does not? Is it mainly used to prevent unstable switching, or does it provide a different basis for deciding when recovery is appropriate? The paper should explain this distinction more directly.

Since an appendix is allowed, why is there no appendix at all? The paper feels underdeveloped in several places. Why not use that space to explain the design choices, how the components work together, and the limitations more fully? There are quite a few interacting mechanisms here, and I do not think their rationale has been discussed in enough depth.

Questions for authors

What is the main insight gained by combining these components in this particular way? When link avoidance and path recovery request conflicting motions, what principle determines the compromise? Why is the global threat index suitable for deciding when to accelerate recovery, and what does event history add beyond the current state? Also, why not include an appendix to explain these design choices and their limitations more fully?

3 comments

We thank the reviewer for identifying the need to separate comparative performance, component contributions and the scope of the safety evidence. We address these points in three responses.

W1 and Q1 Matched task comparison and timing

Review:

Comparison with existing methods. Could the authors provide quantitative comparisons against representative GVF-based, CBF-based, or MPC-based collision-avoidance methods under matched experimental conditions? In particular, comparisons of minimum collision clearance, path-following error, recovery time, success rate, and computational cost would help establish the practical advantages of the proposed framework.

Response: Section 4.5/Table 3 now compares our original-parameter controller, FASM-MPC and three HARMONIOUS settings in 20 paired narrow-frame environments, with the same robot, joint bounds, 15 mm evaluation radius, 10 s duration and 40 ms command updates; body geometry is checked every 1 ms. Completion requires no observed overlap and at least one net turn.

Our controller and FASM-MPC both complete 20/20. Mean progress is 2.583 versus 2.270 turns (+13.8%); FASM-MPC has lower RMS (21.977 versus 31.152 mm) and greater worst clearance (37.276 versus 8.343 mm). HARMONIOUS k2/d30, k6/d30 and k2/d5 complete 10/20, 2/20 and 6/20, respectively. These are control-layer adaptations: FASM-MPC's GPIO observer and HARMONIOUS's perception stack are excluded, and frame-specific directional guidance remains only in our implementation. The comparison therefore evaluates the implemented controllers, with their performance trade-offs.

Recovery is additionally tested in a separate 20-environment common-release protocol at 40 ms updates: mean 5 mm confirmation time is 0.60055 s for ours, 0.09000 s for FASM-MPC (40 mm margin, penalty 10) and 0.65540/0.21695 s for HARMONIOUS k2/d30 and k6/d30, all with 20/20 completion. FASM-MPC is faster in that recovery protocol; our observed external benefit concerns sustained task progress.

The computation column in Table 3 is measured separately on an Intel Core Ultra 7 265KF with MATLAB R2022b (9.13.0.2049777), Windows x64. Each method receives a 0.4 s warm-up and three sequential 10 s replays of its own archived trace in one fixed frame environment (seed 9201), at common 40 ms updates; excluding each replay's first call leaves 747 retained calls per method. Mean elapsed time is 0.689 ms for ours, 184.403 ms for FASM-MPC and 0.951/0.974/0.581 ms for the three HARMONIOUS settings. Our 95th percentile is 1.075 ms, with no retained call exceeding 40 ms.

Timing includes controller geometry, filtering, optimization and internal bookkeeping; it excludes physical integration, plotting, file I/O and independent clearance checking. The 20-environment performance trajectories do not inject computation delay, and these controller-call measurements are separate from end-to-end hardware latency. The full task comparison and mean computation column are in Section 4.5/Table 3; the additional recovery protocol is reported here.

W2 and Q2 Soft coordination after command limiting

Review:

Collision-safety guarantees. Given that Equation (7) formulates link-level avoidance as soft optimization objectives and Equation (10) applies componentwise joint-velocity saturation, under what conditions can collision avoidance be guaranteed? If formal guarantees are not available, could the authors clarify the limitations of the safety claims and provide additional evidence concerning controller behavior under challenging or uncertain conditions?

Response: Equation (7) coordinates link-level avoidance corrections as soft tasks; Eq. (10) explicitly bounds the final joint-velocity commands. Prediction, finite-thickness corrections and threat-dependent weights shape avoidance, and geometric evaluation follows the realized motion after the complete filtering and saturation pipeline.

At 40 ms command updates and 1 ms geometry checks, the original controller completes 20/20 narrow-frame tasks with worst sampled full-capsule clearance of 8.343 mm (Table 3). The controlled body-protection ablation gives 0/20 overlaps for full control versus 20/20 for end-effector-only protection under the same true link geometry. This supports collision-free execution with bounded joint commands in the tested modeled conditions; Section 3.3 makes this empirical scope explicit rather than presenting a general forward-invariance theorem.

A separate archived diagnostic with setting C065, 40 ms updates and 1 ms geometry checks records overlap in 0/20 environments without sensing delay, 1/20 at 80 ms and 2/20 at 120 ms, characterizing sensitivity to delayed information for that setting.

The C065 diagnostic is distinct from the default-controller results and computation timing. Archived noncircular-path checks of the original controller also have no observed overlap in four path-scene combinations with 20 seeds each, totaling 80 run conditions rather than 80 independent environments. A separately parameterized setting C has three overlaps in 20 static-ellipse conditions. These checks characterize the tested conditions and parameter sensitivity; they are not a generalization guarantee. Physical execution evidence is addressed separately below.

W1 and Q3 Component contributions and repeated evaluation

Review:

Component contributions and experimental reliability. Could the authors clarify the number of independent trials or avoidance events used to calculate the success rates in Table 2? In addition, can the individual contributions of predictive activation, capsule-based link correction, and rank-discounted threat aggregation be isolated through ablation experiments?

Response: The additional tests separately examine predictive activation, full-link correction and rank discount. Removing prediction at unchanged thresholds produces 2/20 overlaps versus 0/20 with prediction. Removing applied link correction or modeled thickness gives 20/20 frame overlaps versus 0/20 for full control. The 180-trajectory discount test finds small paired motion changes with unchanged collision/completion outcomes. The full-link and compensation/feedforward controls are reported with our component replies; their roles are not inferred from the original illustration alone.

Each method in the original Table 2 has one static trajectory, with reconstructed completed/released event counts of None 1/1, Fixed window 0/1, State-driven 1/2 and EDSR 1/1.

The extra state-driven event is interrupted by renewed threat, and the fixed-window event is unfinished at the simulation end. These are within-run fractions, with time and integrated error summarized over completed recoveries. Table 2 keeps the original displayed values while clarifying the recovery windows and completed/released counts; the windows are reconciled through numerical reconstruction from archived trajectories.

The repeated common-release test covers 40 environments, with 40/40 recovery for all four strategies. EDSR has mean confirmation time of 0.4478 s versus 0.4447 s for gain-matched state-driven control, a 3.1 ms difference, with approximately 51% lower mean command variation. The persistent-obstacle test covers 20 new environments with four primary strategies (80 trajectories), plus a 20-trajectory intensity-rate diagnostic; it evaluates stable near-path duration rather than an event success probability. These studies distinguish recovery speed and command variation from repeated-trial reliability. The component and repeated-environment results are supplied in the responses; Section 4.2/Table 2 now makes the original event denominator explicit.

W2 and Q3 UR7e execution evidence

Review:

Quantitative measurements and repeated trials on the physical robot would also help assess the reliability of the reported improvements.

Response: The complete controller has been deployed on the UR7e for path following, obstacle avoidance and rejoining, with joint-angle logs recorded. Section 4.4 presents the physical execution snapshots and specifies those recordings. This establishes deployment and physical execution of the model-based controller using known kinematics and explicit link geometry.

The repeated-environment statistics and quantitative path-error/clearance comparisons above are simulation results; the hardware experiment is reported as a qualitative execution demonstration. Reporting these forms of evidence separately keeps their respective roles clear.

Reviewer_6pz5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a controller for manipulator path following with obstacle avoidance that accounts for the geometry of all modeled robot links. The proposed method combines an implicit-surface guiding vector field, approach-only obstacle prediction, moving-boundary compensation, capsule-based link monitoring, and weighted joint-space correction. Overall, it is a competent research paper in the field of robotic controls.

Critical strengths
  1. This paper addresses a practical problem in manipulator path following: intermediate robot links may collide with environmental obstacles even when the end effector follows its desired path. Accounting for collision risks involving all modeled links, rather than focusing solely on the end effector, is practically meaningful.

  2. Although the proposed framework combines several components, their integration is natural and each component serves a clear functional purpose. The framework does not appear to introduce additional techniques, such as LLMs, merely for the sake of incorporating fashionable methods.

Critical weaknesses
  1. This paper primarily proposes a manipulator path-following controller composed of several analytical control components. Its core methodology involves guiding vector fields, robot geometric modeling, and joint-space optimization, without an explicit machine learning component. However, this paper does not provide a clear connection between the proposed method and ML. Therefore, I believe the current work is more closely aligned with traditional robotics control paper than an ICLR paper.

  2. Although thispaper includes an ablation study on EDSR, it lacks systematic ablations of the other key components and their interactions.

Questions for authors
  1. The proposed method mainly involves analytical guiding vector fields, geometric obstacle avoidance, and joint-space control, without an explicit ML contribution. The authors should explain what new methodology or insights this work provides that would be particularly relevant to the ICLR community?

  2. The authors provide more comparisons with existing full-link obstacle avoidance controllers. Moreover, could the authors separately remove the prediction mechanism, full-link correction, and EDSR to evaluate the individual contributions of each components?

  3. Eq. (7) uses weighted least squares to approximate the desired obstacle-avoidance correction velocities for multiple links. However, this optimization aims to find a solution that minimizes the weighted residuals rahter than explicitly enforcing collision-free constraints.

2 comments

We thank the reviewer for the comments on the methodological scope and component evaluation. We clarify the model-based contribution below and provide controlled comparisons in the following response.

W1 and Q1 Methodology and relevance

Review:

The proposed method mainly involves analytical guiding vector fields, geometric obstacle avoidance, and joint-space control, without an explicit ML contribution. The authors should explain what new methodology or insights this work provides that would be particularly relevant to the ICLR community?

Response: The controller is deliberately model-based: known robot kinematics, finite-thickness link geometry and explicit threat/recovery states determine online motion, without policy training. The methodological contribution is the coordination of continuous path guidance, whole-link avoidance corrections and event-conditioned recovery during execution.

The controlled tests make that contribution measurable. Under the same true link geometry, the full-link ablation gives 0/20 overlaps for full control and 20/20 for end-effector-only protection. From common release histories in 40 environments, EDSR reduces mean confirmation time from 0.8804 to 0.4478 s relative to no enhancement; the comparison below further separates recovery speed from command variation. Section 4.5/Table 3 evaluates progress and completion together with tracking error, so small error during stalled motion is not mistaken for successful execution.

These are control and evaluation questions for embodied systems, relevant to the robotics, autonomy and planning topics in the ICLR 2027 call for papers. The execution layer could complement learned high-level policies or planners, but such integration is a potential application rather than a tested result. We position the contribution as model-based online robot control and its empirical evaluation.

W2 and Q2 External comparisons and component contributions

Review:

The authors provide more comparisons with existing full-link obstacle avoidance controllers. Moreover, could the authors separately remove the prediction mechanism, full-link correction, and EDSR to evaluate the individual contributions of each components?

Response: Section 4.5/Table 3 adds a comparison of the original-parameter controller with control-layer adaptations of FASM-MPC and three HARMONIOUS settings in 20 paired narrow-frame environments. The original controller and FASM-MPC both complete 20/20; our controller makes 2.583 versus 2.270 turns of mean progress, while FASM-MPC has lower RMS and greater clearance. The comparison concerns the implemented controllers: it excludes FASM-MPC's GPIO observer and HARMONIOUS's perception stack, and only our implementation retains frame-specific directional guidance.

The additional component tests separately remove prediction, applied full-link correction and recovery enhancement. At unchanged thresholds, removing prediction changes overlaps from 0/20 to 2/20. Removing applied link correction while retaining monitoring changes frame overlaps from 0/20 to 20/20 and mean minimum clearance from 10.620 to -31.779 mm. Compensation and feedforward are examined in a 2 by 2 ablation. Varying the rank discount over 0, 0.35 and 1 in 180 trajectories gives small paired motion changes with unchanged collision/completion outcomes. These tests isolate selected mechanisms; they are not a full factorial of all component interactions.

For recovery, four strategies start from common pre-release histories in 40 environments; all recover in 40/40, with confirmation defined by a 5 mm criterion.

Response Table R6. Common-release recovery comparison. RMS and command variation use the same 4 to 8 s interval after obstacle withdrawal at 4 s. Command variation sums squared vector increments between consecutive applied joint-velocity commands, in (rad/s)2(\mathrm{rad}/\mathrm{s})^2; lower is better.

StrategyConfirmation time (s)RMS (mm)Command variation
Full EDSR0.44784.5530.03117
No enhancement0.88045.9480.02964
State-driven0.43704.4300.06711
Gain-matched state-driven0.44474.4580.06394

Compared with gain-matched state-driven control, EDSR has similar mean confirmation time, differing by 3.1 ms, and approximately 51% lower mean command variation. Relative to no enhancement, mean confirmation time is 49.1% shorter. This supports a recovery-speed and command-variation trade-off, rather than a higher recovery success rate. The gain-matched control here differs from the auxiliary intensity-rate diagnostic in the separate persistent-obstacle experiment. Section 4.5/Table 3 is added to the manuscript; these component and repeated-environment results are supplied in the responses.

Q3 Soft coordination and bounded joint commands

Review:

Eq. (7) uses weighted least squares to approximate the desired obstacle-avoidance correction velocities for multiple links. However, this optimization aims to find a solution that minimizes the weighted residuals rather than explicitly enforcing collision-free constraints.

Response: Equation (7) coordinates soft avoidance tasks, while Eq. (10) explicitly enforces the configured joint-velocity bounds. Prediction, finite-thickness normal corrections and threat-dependent weights act before the final command is filtered and clipped; clearance is then evaluated on the motion generated by those final commands.

With 40 ms updates and 1 ms geometry checks, the original controller completes 20/20 frame tasks with worst sampled full-capsule clearance of 8.343 mm (Section 4.5/Table 3). Under the same true link geometry, full body protection also gives 0/20 overlaps versus 20/20 for end-effector-only protection. These results support collision-free motion with bounded joint commands in the evaluated modeled conditions. Section 3.3 clarifies that this empirical result is distinct from a general forward-invariance theorem.

Reviewer_nn3XOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

A six-DOF arm has to follow a path while every modelled link, not just the end effector, keeps away from obstacles, and it has to rejoin the path afterwards. Eq. (3) predicts how close the obstacle will come over a horizon T_psi, and Eq. (5) picks a danger point on each capsule link so that link thickness can be folded into the obstacle level. A rank-discounted product (Eq. 6) merges the per-link threats into one number, which weighted damped least squares turns into joint velocities (Eq. 7). Rejoining (Eq. 8 and 9) switches on only when that number drops. Results are three MATLAB scenarios, a CoppeliaSim co-simulation and a UR7e.

Critical strengths
  1. The method is laid out equation by equation, from the path field through the capsule link model (Eq. 5) to the joint-velocity step, so the block structure is easy to follow and to reuse.

  2. The narrow-frame experiment (Fig. 2) is the paper's best evidence: the end-effector level stays outside the reactive shell while up to three links are activated, at 0.0107 m clearance, and the static case shows the prediction entering the shell 0.351 s early.

Critical weaknesses
  1. The ablation does not support the safety claim in the abstract. Its None row already reports 100% success and wider clearance than EDSR, 0.0503 against 0.0493 m, so what the method gains is speed and less accumulated error: Trejoin 0.823 to 0.471 s, Jrejoin 8.56e-3 to 4.90e-3. The four rows also keep 4.87 to 5.09 cm of clearance, against 1.07 cm in the narrow frame, so the ablation ran where safety was never at risk.

  2. The method is never compared with an existing approach. All four rows of Table 2 are the authors' own variants. Section 2 says conventional blending generally returns to the nominal field once clearance is regained, without controlling the recovery, yet none of the methods cited there is run, including the CBF-guided MPC of Liu et al. (2025).

  3. Two reasons the paper gives for the design are stated but never measured. In the moving test the actual obstacle level never comes inside the warning zone, lowest 2.589 against 2.40, so avoidance is driven by the prediction alone, and vcomp and vff are never tested on their own. Section 4.1 says the narrow-frame run shows that watching only the end effector is not enough, but that version is never run.

Questions for authors
  1. The no-rejoining variant already reaches 100% success with wider clearance than EDSR, 0.0503 against 0.0493 m. Is there any case where not rejoining ends in failure?

  2. Table 2 compares only your own variants, while Section 2 implies the method handles recovery better than conventional blending. Would one external run, such as the CBF-guided MPC you cite, support that?

  3. In the narrow frame the end-effector level stays outside the reactive shell, so a controller that guards only the end effector would never react. What clearance does that version reach?

  4. In the moving test the actual obstacle level never comes inside the warning zone, closest 2.589 against 2.40. What changes if vcomp and vff are removed?

  5. Table 1 omits rho_on, rho_off, k_ff and k_pre. Since rho_on decides which links are treated as active, the reported link counts cannot be checked.

  6. Table 2 reports 4.87 to 5.09 cm clearance, while Section 4.1 reports 3.20, 2.54 and 1.07 cm, and Section 4.2 names no scenario. Which setting produced Table 2, and how long was the fixed window?

  7. Section 4.4 gives six photographs and no numbers. What minimum clearance and path error did the UR7e run reach?

  8. Eq. (6) fixes q_Gamma at 0.35 and no other value is tried, yet it decides how much the second and third most dangerous links count. How do the results change when it is varied?

6 comments

We thank the reviewer for the detailed assessment. We address recovery, body protection, external comparisons and reproducibility separately below, using the completed controlled tests and the focused manuscript revision.

W1 and Q1 Recovery efficiency and recovery before the next encounter

Review:

W1: "The ablation does not support the safety claim in the abstract."

Q1: "The no-rejoining variant already reaches 100% success with wider clearance than EDSR, 0.0503 against 0.0493 m. Is there any case where not rejoining ends in failure?"

Response: The original Table 2 supports a recovery-efficiency benefit: EDSR reduces rejoining time from 0.823 to 0.471 s and integrated rejoining error from 8.56×10−38.56\times10^{-3} to 4.90×10−34.90\times10^{-3} m s. None retains slightly larger clearance, 50.3 versus 49.3 mm, and both complete their single released event. The lower error integral also reflects the shorter recovery window. We therefore distinguish this recovery result from the body-protection evidence reported in Response 3 below.

A moving-obstacle example shows why recovery speed matters before another encounter. Renewed threat interrupts all five unenhanced recovery episodes, whereas EDSR completes 4/4. At the first common release, both have 60.148 mm path error; EDSR completes at 2.079 s, while None still has 8.639 mm error when threat returns at 2.721 s. Both keep positive clearance. This is failure to finish recovery before the next encounter, not collision or permanent nonconvergence; the later events depend on the resulting trajectories and are not independent trials.

We also evaluate stable near-path operation in 20 new persistent-obstacle environments:

Response Table R1. Stable near-path time in 20 persistent-obstacle environments. The duration sums complete segments with error at most 5 mm lasting at least 50 ms, including the first 50 ms. Values are means across environments.

StrategyStable near-path time (s) ↑
Full EDSR1.98590
No enhancement0.23250
Development-selected fixed window1.40595
State-driven2.02105

EDSR adds 1.7534 s relative to no enhancement, with paired 95% bootstrap interval [1.6861, 1.8149] s. The state-driven strategy attains slightly longer duration. An additional diagnostic with the same ordinary intensity-rate limits gives 1.99455 s, close to EDSR's 1.98590 s. At the stricter 3 mm tolerance, the fixed window exceeds EDSR, 1.04525 versus 0.60445 s; the 3/5/10 mm evaluations reuse the same trajectories. These results support the task value of recovery enhancement while keeping the conclusion specific to its measured effects.

The separate 40-environment common-release test assesses recovery speed and command variation: EDSR takes 0.4478 s versus 0.4447 s for gain-matched state control, while command variation is approximately 51% lower, 0.03117 versus 0.06394; all four strategies recover in 40/40. This is a trade-off between recovery speed and command variation, rather than higher recovery success probability. Section 4.2 and manuscript Table 2 now clarify the original metric windows and within-run event counts.

W2 and Q2 Comparisons with existing controllers

Review:

W2: "The method is never compared with an existing approach. All four rows of Table 2 are the authors' own variants."

Q2: "Table 2 compares only your own variants, while Section 2 implies the method handles recovery better than conventional blending. Would one external run, such as the CBF-guided MPC you cite, support that?"

Response: Section 4.5 and Table 3 now compare the original-parameter controller with control-layer adaptations of FASM-MPC and HARMONIOUS. All use the same 20 paired narrow-frame environments, robot, joint bounds, true 15 mm-radius evaluation capsules, 10 s horizon and 40 ms command updates, with geometry checked at 1 ms intervals. Completion requires no observed overlap and at least one net turn.

Response Table R2. Results reproduced from manuscript Table 3. RMS and progress are means; clearance is the worst sampled minimum among the 20 environments. Computation is measured separately in three 10 s replays of one fixed frame environment (seed 9201), also at 40 ms updates.

ControllerCompleted /20RMS (mm)Progress (turns)Worst clearance (mm)Mean computation (ms)
Ours, original parameters2031.1522.5838.3430.689
FASM-MPC2021.9772.27037.276184.403
HARMONIOUS, k2/d301024.6321.67229.9260.951
HARMONIOUS, k6/d30218.4240.49127.2530.974
HARMONIOUS, k2/d5620.3221.1404.6190.581

Our controller and FASM-MPC complete 20/20; ours achieves 2.583 versus 2.270 mean turns, a 13.8% increase, with lower measured computation per update. FASM-MPC has lower RMS and larger clearance. The timing column is a separate nominal-environment measurement; performance trajectories do not inject computation delay. It characterizes controller-call cost, not end-to-end hardware latency.

The adaptations omit FASM-MPC's GPIO observer and HARMONIOUS's perception stack, while our controller retains frame-specific directional guidance absent from the baselines. The comparison therefore concerns these implemented controllers; it does not isolate EDSR or reproduce either external paper's complete system.

We additionally compare recovery from a common release history generated by the original controller in 20 environments at 40 ms updates. Mean 5 mm confirmation time is 0.60055 s for ours, 0.09000 s for FASM-MPC (40 mm margin, penalty 10), and 0.65540/0.21695 s for HARMONIOUS k2/d30 and k6/d30; all recover in 20/20. FASM-MPC is faster in that protocol. Our measured external benefit is sustained task progress, while the within-framework comparison identifies the effect of recovery enhancement. Manuscript Section 4.5/Table 3 contains the task and computation comparison; this independent recovery protocol is supplied in the response.

W3 and Q3 Full-link monitoring and finite link thickness

Review:

W3: "Section 4.1 says the narrow-frame run shows that watching only the end effector is not enough, but that version is never run."

Q3: "In the narrow frame the end-effector level stays outside the reactive shell, so a controller that guards only the end effector would never react. What clearance does that version reach?"

Response: The controlled end-effector-only comparison is now available. In 20 paired narrow-frame environments, all versions are evaluated using the same six true 15 mm-radius link capsules, even when the controller protects only the end effector. Prediction, EDSR and the other settings are retained.

Response Table R3. Full-link ablation. Mean minimum clearance first takes the whole-run minimum in each environment, then averages the 20 minima. A negative value denotes geometric overlap.

ControllerMean minimum clearance (mm)Environments with overlap
Full controller10.6200/20
Applied link correction disabled-31.77920/20
End-effector protection only-31.77920/20
Zero modeled link thickness-4.26420/20

End-effector-only protection gives mean minimum body clearance of -31.779 mm and overlap in 20/20, compared with 10.620 mm and 0/20 for full control. The separate applied-correction and zero-thickness ablations also overlap in 20/20. Thus body geometry and applied link correction account for the protection observed in these conditions. Lower tracking error or greater progress while the body overlaps an obstacle is not successful execution.

Section 4.1 now distinguishes the original motivating illustration from this direct controlled comparison. The additional ablation is reported here; it is not added as another manuscript subsection.

W3 and Q4 Prediction, compensation and obstacle-velocity feedforward

Review:

W3: "In the moving test the actual obstacle level never comes inside the warning zone, lowest 2.589 against 2.40, so avoidance is driven by the prediction alone, and vcomp and vff are never tested on their own."

Q4: "In the moving test the actual obstacle level never comes inside the warning zone, closest 2.589 against 2.40. What changes if vcomp and vff are removed?"

Response: Prediction concerns when avoidance activates, whereas compensation and feedforward determine the activated motion. The current level remaining above threshold, 2.589 versus 2.40, therefore does not make the latter terms inactive. In 20 paired moving-obstacle environments, removing prediction at unchanged thresholds changes overlap from 0/20 to 2/20; enlarging the reactive thresholds restores positive clearance with tracking/progress trade-offs.

The feedforward term vff=kffηvobs\boldsymbol v_{\mathrm{ff}}=k_{\mathrm{ff}}\eta\boldsymbol v_{\mathrm{obs}} accounts for obstacle translation. It is subtracted before shaping the avoidance field in the relative-velocity frame and added back afterward. Normal compensation supplies the outward motion required by an approaching obstacle boundary. We retain prediction and the other settings in the following 2 by 2 ablation:

Response Table R4. Compensation and feedforward ablation. Clearance is the mean of whole-run minima; overlap counts environments.

Compensation groupFeedforwardOverlap /20Mean minimum clearance (mm)RMS (mm)
OnOn029.97643.637
OffOn121.27036.607
OnOff030.01843.494
OffOff121.79639.031

The compensation switch removes the implemented end-effector/link normal-compensation group, including boundary-motion compensation and cancellation of approaching background normal motion; it does not isolate only the boundary-speed subterm. Removing that group lowers mean minimum clearance and introduces one overlap. With compensation disabled, retaining feedforward reduces RMS from 39.031 to 36.607 mm; with compensation enabled, both feedforward configurations retain 0/20 overlaps and approximately 30 mm mean minimum clearance. These results separate their physical roles and show how the measured effect depends on configuration. They are response evidence, while the original numbered method equations are retained.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.