acceptodds
Withdrawn from ICLR 2027

GeoPhase: Phase-Aware Geometric Safety Filtering for Vision-Language-Action Policies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language-action (VLA) policies enable general-purpose robot manipulation, but their task-directed actions lack explicit geometric safety guarantees, leaving the robot vulnerable to physical collisions in cluttered environments. While runtime safety filtering via control barrier function quadratic programs (CBF-QPs) offers an interpretable solution, current approaches are bottlenecked by costly multi-stage geometric perception, phase-agnostic constraint scheduling, and insufficient protection scopes that typically ignore the grasped object and arm body. To address these limitations, we present GeoPhase, a runtime geometric safety layer that operates on a hybrid learned-perception and analytic-control design. GeoPhase bypasses explicit perception pipelines by predicting structured obstacle ellipsoids directly from intermediate VLA visual embeddings. By tracking the manipulation phase (approach, grasp, transport, place) online via proprioceptive and gripper signals, it dynamically adapts constraint activation to optimize the phase-specific balance between safety and task completion. Furthermore, GeoPhase constructs a multi-proxy geometric representation of the robot-object system, encompassing the end-effector, held object, and arm body, to jointly constrain the robot–object geometry via a unified, minimal-deviation CBF-QP action projection. On the SafeLIBERO benchmark, GeoPhase achieves an average collision rate of and a safe success rate of , outperforming the best reported baseline averages by and percentage points, respectively. These results demonstrate that GeoPhase substantially improves collision-free task completion without modifying the underlying, frozen VLA policy.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Visual Language Action Models, Safe Robotic Manipulation, Control Barrier Function

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

The paper studies runtime safety filtering for VLA-based robotic manipulation (in single-obstacle settings). Existing CBF-based approaches rely on costly multi-stage geometric perception and restrict protection to the end-effector. To address these, the authors train a lightweight geometry prediction head with supervised labels to estimate the obstacle’s geometry from frozen VLA visual representations. At inference, a manually defined finite-state machine identifies manipulation phases (approach, grasp, transport, and placement) and adjusts CBF constraints. To extend collision protection beyond the end-effector, the method models robot body links as proxies whose motions are estimated through analytical kinematic Jacobians, while bounding the grasped object with a conservative envelope; both are modeled as ellipsoids. A CBF-QP then minimally modifies the VLA’s Cartesian actions to satisfy the active geometric safety constraints without retraining the policy. Experiments in simulation and on real robots demonstrate improved collision avoidance and safe task completion.

Critical strengths
  • The method addresses the practical challenge of improving VLA safety without retraining. I really like the design that preserves the VLA’s Cartesian action space by using analytical kinematic Jacobians to relate Cartesian commands to individual robot-link velocities. This extends collision protection beyond the end-effector while maintaining compatibility with the Cartesian action interfaces commonly used in VLA-based manipulation.

  • The experiments include ablation studies and real-world robotic evaluations.

Critical weaknesses
  • The geometry prediction head appears to be the major contributor to performance gains, but its supervision requirements are insufficiently characterized. In Table 2, replacing the learned geometry head with AEGIS’s perception pipeline yields performance comparable to that of AEGIS in Table 1. This does not negate the contributions of the other components, but suggests that their effectiveness depends strongly on accurate obstacle geometry, which is intuitive. A key question is whether collecting geometric supervision is substantially less burdensome than collecting additional safe demonstrations for direct VLA fine-tuning. Although geometric labels can be readily derived from simulator collision meshes, obtaining equivalent supervision in real-world settings may be considerably more challenging. The authors report roughly 63K supervised frames in simulation and mention manually annotated physical data for real-world deployment, but provide limited information about the latter’s size, collection procedure, or annotation effort.

    • The geometry prediction head relies on standard supervised learning. This may raise potential coverage concerns about the supervision data and is conceptually straightforward.
  • While the proposed whole-body safety filtering is interesting, it relies on how the low-level controller resolves the ambiguity in robot-link motions for a given end-effector command. The authors explicitly assume negligible null-space motion in the appendix, but this is an important assumption that should be highlighted in the main paper.

  • The phase-aware scheduling is heuristic and the contribution (Table 2) seems modest.

Questions for authors
  • What are the data collection and annotation procedures for the real-world geometry prediction head including annotation effort?

  • What is the measured latency of AEGIS’s geometry perception pipeline? AEGIS reports a control frequency of 20 Hz, while GeoPhase achieves approximately 21.4 Hz despite enforcing additional safety constraints for robot body links and the held object.

3 comments
AuthorsOpenReview ↗

Response-1

We sincerely thank Reviewer JBDp for the careful reading and constructive assessment. We are encouraged that you recognized our use of analytical kinematic Jacobians to extend collision protection beyond the EEF while preserving compatibility with Cartesian action interfaces. We address each concern below and specify the corresponding clarifications and revisions.

W1 & Q1: Real-World Geometry Supervision

We agree that accurate obstacle geometry is important to the downstream CBF-QP and that the real-world supervision procedure should be clearer. Our physical dataset contains approximately 15k control-step samples from 25 static training configurations, separate from evaluation, with approximately 5 hours of manual geometry annotation. The observations are collected through rollouts of the frozen π0.5\pi_{0.5} policy under varied obstacle placements and scene layouts.

We first calibrate the tabletop frame to the robot base frame. For each obstacle, we manually measure its bounding dimensions (L,W,H)(L,W,H) and bounding-box center cTc_T in the tabletop frame. Given the calibrated transformation (RBT,tBT)(R_{BT},t_{BT}), we compute cB=RBTcT+tBTc_B=R_{BT}c_T+t_{BT} and the base-frame bounding-box half-extents bB=∣RBT∣(L,W,H)⊤/2b_B=\lvert R_{BT}\rvert(L,W,H)^\top/2. We set the ellipsoid semi-axes to rB=3,bBr_B=\sqrt{3}\\,b_B, which conservatively encloses the measured bounding box. For a fixed obstacle pose, the resulting label is reused across all frames in the rollout. Translating the obstacle requires updating only its center; changing its orientation requires recomputing the base-frame extents. Thus, the supervision is amortized across frames and does not require frame-by-frame manual annotation or additional collision-free action demonstrations.

ItemPhysical dataset
Training configurations25 static configurations, 5 per task template, separate from evaluation
Supervised samples~ 15,000 control-step samples
Training objectsBowl, Cola, cup, kettle, and milk
Unseen evaluation instancesWine bottle and the additional mug and beverage bottle reported in W1.1
Manual measurement and annotation~ 5 hours

We will add the collection procedure, label construction, and dataset statistics to Appendices C.1 and E.6.

W1.1: Standard Supervised Geometry Head and Coverage

A: We agree that the geometry head uses standard supervised learning. Our contribution is a structured geometric interface between frozen VLA representations and analytic safety filtering. The supervision combines ellipsoid-parameter regression with barrier-value and barrier-gradient losses, with additional weight near the safety boundary.

For simulation, the geometry dataset contains 63,187 frame-level samples and is split strictly at the rollout level into training and validation sets. Evaluation rollouts are excluded from training and model selection. These experiments assess generalization across spatial configurations and trajectories within a defined operational domain.

In the physical experiment, the unseen wine bottle provides evidence of transfer to one held-out obstacle instance under the fixed-camera setup, with an approximately 2,cm2\\,\mathrm{cm} center error. We also evaluated two additional unseen physical obstacles (10 trials each): an OOD mug(≈1 cm\approx 1\text{ cm} center error) and an OOD beverage bottle (≈2 cm\approx 2\text{ cm} center error). We will include these expanded OOD evaluations in Appendix E.6.

W2: Low-Level Controller and Null-Space Motion

A: We agree that the negligible null-space motion assumption should be stated in the main paper. Our link-motion model uses Tℓ(i)←e=Jℓ(i)Je†T_{\ell(i)\leftarrow e}=J_{\ell(i)}J_e^\dagger, which assumes negligible auxiliary null-space velocity, (I−Je†Je)ν≈0(I-J_e^\dagger J_e)\nu\approx0.

We will state this assumption explicitly in Section 4.3 and clarify that the body-proxy constraints are conditional on the deployed controller satisfying this approximation.

AuthorsOpenReview ↗

Response-2

W3: Contribution of Phase-Aware Scheduling

A: We agree that the FSM-based phase tracker is heuristic. However, this lightweight design is deliberate, and its contribution to balancing safety and task completion is substantial.

Different manipulation phases exhibit distinct geometric risks and control requirements (Sec. 4.2): approach and grasp require sufficient proximity to the target, transport introduces the held object into the protected geometry, and place requires precise release. Unlike the phase-agnostic AEGIS and KNOWS, GeoPhase’s online FSM dynamically modulates the active proxy set PσtP_{\sigma_t}, activation threshold hfree(σt)h_{\text{free}}(\sigma_t), and barrier gain αi(σt)\alpha_i(\sigma_t). It aims to adapt safety parameters to manipulation phases using gripper commands and proprioceptive history, without task-specific phase annotations.

The phase tracker costs 0.03 ms per control step (Table 7) while supporting multi-cycle manipulation such as the SafeLIBERO LONG suite. As shown in Table 2 and Figure 4(b), phase awareness increases overall TSR from 63.8% to 66.4% and SSR from 59.4% to 62.6%, while reducing CR from 16.2% to 12.8%.

Q2: Latency Comparison

A: AEGIS’s explicit perception pipeline involves a costly multi-stage process (remote VLM semantic identification →\rightarrow GroundingDINO →\rightarrow depth point-cloud filtering →\rightarrow ellipsoid fitting). Because it relies on remote API calls with high network latency and computationally heavy local point-cloud optimization, a single complete perception update takes seconds.

We evaluated the AEGIS visual pipeline on object_I Task2 (milk), episodes 0--9, using one RTX 4090 (24 GB), dual Xeon Silver 4314 CPUs, and two 1024×10241024\times1024 views. After excluding the first cold-start run, the remaining nine runs yielded:

AEGIS Perception StageMean (ms)P95 (ms)
GroundingDINO + Point-Cloud Projection & Filtering288.08335.59
Ellipsoid Fitting (ConvexHull + CVXPY/SCS)12,019.3223,383.34
Complete Local Vision Pipeline (w/o remote API)14,804.9726,561.75
Full Pipeline (incl. Remote VLM API & retries)40,007.0587,731.21

As shown above, ellipsoid fitting accounts for approximately ~8080\\% of the runtime. These measurements show that the complete pipeline cannot be recomputed within a 50,ms50\\,\mathrm{ms} control step under our setup. They do not establish how the originally reported 20 Hz AEGIS controller schedules or reuses geometry.

In contrast, GeoPhase reuses pooled VLA features and predicts ellipsoid parameters through a lightweight MLP. VLA inference (14.41 ms) and geometry prediction (12.63 ms) run asynchronously per action chunk, while our local safety layer runs synchronously in just 10.21 ms/step, including phase tracking, proxy updates, constraint construction, QP solving, and post-processing. The measured mean end-to-end control computation is 46.90 ms/step (21.4 Hz).

We will add the complete profiling protocol and latency breakdown to Appendix E.2.


To sum up:

We clarified the real-world geometry-collection procedure, annotation effort, and frame-level label reuse (W1&Q1). We further acknowledged the standard supervised nature of the geometry head and expanded the held-out obstacle evaluation (W1.1). We will explicitly state the negligible null-space motion assumption in the revised manuscript (W2). We explained the role and low computational cost of the phase tracker while reporting its scenario-dependent benefit (W3). Finally, we provided a latency breakdown for the AEGIS perception pipeline (Q2).

We sincerely thank the reviewer for the constructive feedback, which has helped us improve the precision and completeness of our presentation.

If you have any further questions or additional concerns, please feel free to discuss them with us.

Dear Reviewer JBDp,

We truly value all of your suggestions. If you have any further questions, please do not hesitate to let us know. We will do our best to address them during the remainder of the discussion period. Thank you again for your time and effort. Best of luck!

Best regards, Authors

Reviewer_JY75OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a method called GeoPhase which is a runtime safety layer which could be used on top of vision language action policies. This approach leverages an MLP that takes as input the visual embedding of the VLA, has been trained to regress obstacle ellipsoids, a finite-state machine that tracks the phase of manipulation, and geometric proxies for the end effector, held object and the robot links used. Phase dependent control barrier function constraints are encoded as a quadratic program that modifies the Cartesian actions of the policy. The papers main contribution is centered around integrating learned geometric perception, constrain scheduling and a proxy safety filter. Experiments are presented on SafeLIBERO and five real world manipulation tasks showing the decrease of collision rate and the improvement of safe task execution.

Critical strengths
  1. The approach considers potential collisions involving the held object and the selected robot links, in addition to the end effector, a practical consideration as these bodies can sometimes collide with obstacles even when the end effector does not.
  2. The paper also assesses the approach on 1600 of SafeLIBERO episodes and five real tasks involving manipulation, with component ablations analyzing the contribution of the different elements. The physical experiments show a decrease of the number of collisions and an improvement of the safe success, evidencing practical advantages.
Critical weaknesses
  1. The contribution in this paper mainly seems to be a combination of some core ideas, rather than a new safety scheme per se. Core ideas such as CBF-based action projection, ellipsoidal collision representations, and supporting-hyperplane barriers are already established, as seen in AEGIS [1]. GeoPhase combines supervised obstacle prediction based on VLA embeddings, a human-designed phase tracker, and geometric proxies. While this combination can be of practical use, more evidence is needed to establish its advantages over simpler alternatives.
  2. The approach activates constraints using phase-dependent proximity thresholds and reports proxy intervention rates. However, these are not sufficient to determine how often corrections prevent collisions, unnecessarily affect safe behaviour, or cause task failures. Additionally, testing on unobstructed and low-risk scenes, alongside the safety-critical benchmark, could help quantify this trade-off.
  3. The static ablation keeps gripper based held-object activation and uses the transport-phase barrier gain throughout the task. However the paper does not compare against a static filter whose gain and activation threshold are tuned together, or a simpler schedule based only on whether the robot is holding an object.
  4. The real world experiments compare Geo Phase with only the unfiltered policy. They demonstrate practical advantages but don't illustrate how well Geo Phase compares to another safety filter on hardware. A comparison with another method, for example, with AEGIS [1], under the same policy and similar tuning budgets, would reinforce these empirical results.

References

  1. Hu, Songqiao, et al. "Vlsa: Vision-language-action models with plug-and-play safety constraint layer." arXiv preprint arXiv:2512.11891 (2025).
Questions for authors
  1. It is not clearn from the paper how far the geometry predictor might be able to generalise to unseen objects with different camera viewpoints as well as to partial occlusion of objects outside of the context of the wine bottle. This raises the question of whether the barrier-aware loss function is superior to the familiar geometry regression technique.
  2. How would the method efficiently cope with several obstacles? Also what happens if the object being held is larger than the fixed proxy?
  3. It is also unclear whether the filter meets its control deadlines. How old were the geometry estimates when used? It would also be interesting to include the missing control settings and give a more detailed description of the real-world training data.
4 comments
AuthorsOpenReview ↗

Response-1

We sincerely thank the reviewer JY75 for the careful reading and constructive feedback. We appreciate your recognition of the practical benefits demonstrated by our multi-proxy protection. Below, we provide detailed responses to address your concerns.

W1: Technical contributions and advantages

A: We agree that CBF-based action projection, ellipsoidal representations, and supporting-hyperplane barriers are established primitives (e.g., Funada et al., 2024; AEGIS). However, GeoPhase's contribution is architecting a tightly coupled runtime safety interface that resolves three system-level bottlenecks left open by prior methods:

  1. Direct Obstacle Geometry Prediction from VLA Embeddings. GeoPhase maps pooled visual embeddings ZtZ_t to obstacle parameters (ct,rt)(c_t,r_t) through a lightweight MLP, avoiding explicit detection, depth processing, and ellipsoid fitting at deployment. The amortized cost is 12.63,ms12.63\\,\mathrm{ms} per control step. Within the same GeoPhase controller, replacing AEGIS geometry with our geometry reduces CR from 34.4% to 12.8% (Fig. 4a). This isolates the contribution of the geometry interface.

  2. Phase-Aware Scheduling. Different manipulation stages exhibit distinct geometric risks and control requirements. The FSM jointly schedules the eligible proxy set, activation threshold, and barrier gain according to the manipulation phase. The validation-tuned static baseline achieves SSR/CR of 59.6%/15.1% (See W3), compared with 62.6%/12.8% for phase-aware scheduling, demonstrating gains beyond fixed-parameter tuning. At the grasp-to-transport transition, GeoPhase additionally verifies the barrier initialization condition before activating the held-object proxy (Appendices C.2–C.3).

  3. Multi-Proxy Control through a Shared Cartesian Action. Unlike AEGIS and KNOWS focus only on EEF–obstacle separation, GeoPhase jointly represents the EEF, the held object, and body links. Because a frozen VLA outputs EEF commands rather than independent link controls, these proxies must be coupled explicitly. Appendix B.1 (Eqs. 13–22) derives the map Tℓ(i)←e=Jℓ(i)Je†T_{\ell(i)\leftarrow e}=J_{\ell(i)}J_e^\dagger and the rigid-offset coupling matrix SiS_ito express each proxy’s motion and separating-direction update within one affine constraint system. Thus, one QP can coordinate the shared Cartesian action across the EEF, held-object, and body proxies.

These components have complementary benefits: replacing learned geometry, disabling phase scheduling, or retaining only the EEF proxy reduces SSR from 62.6% to 45.6%, 59.4%, and 56.2% (Table 2 and Fig. 4). The full-system comparison in Table 1 gives GeoPhase 12.8% CR and 62.6% SSR, while 37.6% and 45.4% for AEGIS.

W2: Intervention effectiveness and performance in low-risk scenes

A: We agree that intervention rates do not determine whether a correction prevents a collision, unnecessarily changes safe behavior, or causes task failure. Table 4 characterizes proxy-specific intervention patterns. We complement this analysis with episode outcomes and an additional low-risk evaluation.

  1. Safety-critical Evaluation. The outcome decomposition in Appendix E.1 shows that success with collision decreases from 40.6% to 3.8%, and failure with collision from 41.7% to 9.0%. Meanwhile, SSR increases from 15.8% to 62.6%, TSR from 56.4% to 66.4%, and mean ETS from 282.4 to 264.3. These results confirm that interventions primarily prevent collision-induced failures without causing excessive detours.

  2. Evaluation without obstacles. We additionally evaluated the same 16 original LIBERO tasks without the obstacles introduced by SafeLIBERO, using 10 episodes per task and 160 episodes total. Aligned with SafeLIBERO, a collision is recorded if task-irrelevant objects are displaced by dt=∑j∥pj,t−pj,0∥2>0.001,md_t=\sum_j\lVert p_{j,t}-p_{j,0}\rVert_2 > 0.001\\,\mathrm{m} | Suite | Base TSR | GeoPhase TSR | Base CR | GeoPhase CR | Base ETS | GeoPhase ETS | GeoPhase IR | |:--|:--:|:--:|:--:|:--:|:--:|:--:|:--:| | Spatial | 97.5 | 97.5 | 2.5 | 2.5 | 262.0 | 265.5 | 3.0 | | Goal | 95.0 | 92.5 | 2.5 | 0.0 | 251.5 | 254.5 | 2.8 | | Object | 97.5 | 95.0 | 5.0 | 2.5 | 278.5 | 283.8 | 4.5 | | Long | 87.5 | 90.0 | 10.0 | 2.5 | 321.5 | 328.0 | 6.0 | | Overall | 94.4 | 93.8 | 5.0 | 1.9 | 278.4 | 283.0 | 4.1 |

GeoPhase intervenes on only 4.1% of steps, preserving task success and execution efficiency, while reducing incidental background contacts to 1.9% CR. The filter's design supports this behavior. If every eligible pair satisfies hi,t≥hfree(σt)h_{i,t}\ge h_{\mathrm{free}}(\sigma_t), then Kt=∅\mathcal{K}_t=\emptyset and the nominal command is executed unchanged. If constraints are active but the nominal augmented decision vector satisfies all of them, the projection also returns the nominal solution ut∗=ut0u_t^* = u_t^0.

We will add these results in the revision.

AuthorsOpenReview ↗

Response-2

W3: Comparison with Tuned Static Baseline

A: We clarify that our original static baseline (Section 5.3) already activates the held-object proxy according to holding status, while using a shared gain and activation threshold throughout the task. Thus, it instantiates a simpler binary schedule based solely on whether the robot is holding an object.

To eliminate any tuning bias, we evaluated a global static baseline on the same validation set and with a matched tuning protocol. We searched αstatic∈�0�s−1,hfree,static∈�1�m.\alpha_{static}\in�0�s^{-1},h_{free,static}\in�1�m.

αstatic\alpha_{static} / hfree,statich_{free,static}0.030.060.10
2.554.7 / 21.656.1 / 20.357.0 / 19.5
557.6 / 19.459.2 / 17.860.1 / 16.4
1058.8 / 17.460.9 / 15.561.9 / 14.1
1557.9 / 18.260.0 / 16.261.4 / 14.3
2055.6 / 20.158.3 / 18.059.5 / 16.5

We then evaluated the optimal configuration (αstatic=10s−1,hfree,static=0.10m\alpha_{static} = 10 s^{-1}, h_{free, static} = 0.10 m) on the full 1,600-episode SafeLIBERO test set:

Methodα\alpha (s−1s^{-1})hfreeh_{free} (m)SSR ↑CR ↓
Original static baseline150.1059.416.2
Tuned global static100.1059.615.1
Phase-aware GeoPhase(10,10,15,5)(0.10,0.06,0.10,0.03)62.612.8

This confirms that the improvement arises from phase-aware scheduling rather than tuning bias: a single static configuration cannot simultaneously accommodate the permissive proximity required in Grasp/Place and the expanded protective envelope needed during Transport.

We will incorporate these results into revised manuscript.

W4: Real-world comparison with AEGIS

A: Following the reviewer's suggestion, we deployed AEGIS on the same Franka Research 3 and five manipulation tasks, with 10 trials/task. The comparison uses the same π0.5\pi_{0.5} checkpoint, observation configuration, tuning budget, and success/collision criteria.

TaskBase TSR ↑CR ↓SSR ↑AEGIS TSR ↑CR ↓SSR ↑GeoPhase TSR ↑CR ↓SSR ↑
1706010703050701060
2907030903060901080
3807010804050902070
4806020803060801070
58060209030601001090
Avg.806418823256861274

GeoPhase reduces CR by 20 pp and increases SSR by 18 pp relative to AEGIS, with TSR increasing from 82 to 86. A representative Task 5 rollout illustrates the benefit of the broader protection scope: AEGIS redirects the EEF around the kettle, but an unprotected arm link subsequently contacts it. GeoPhase includes arm-link and held-object proxies in the shared action projection, addressing this collision mode. This observation is consistent with the multi-proxy ablation in simulation.

We will add the comparison to revised manuscript.

Q1: Geometry Generalization & Barrier-Aware Loss

A: We distinguish generalization within the evaluated camera setup from generalization to changed camera viewpoints, and provide additional unseen-object and loss-ablation results.

  1. Occlusion and Viewpoints. The 90%/10% split in Appendix C.1 is a rollout-level training/validation split; evaluation rollouts are separate from both. Evaluation includes randomized layouts and partial self-occlusion. The reported coverage of 0.931 and IoU of 0.846 aggregate all evaluated frames, including partially occluded observations. The physical predictor is trained with a fixed camera configuration and base-frame supervision; changing the camera perspective requires reannotation and retraining.

  2. Unseen Objects. Beyond the wine bottle, we evaluated two additional obstacle instances excluded from geometry-head training, with 10 trials:

ObstacleCenter ErrorBase CR ↓GeoPhase CR ↓
Wine bottle~2 cm6010
Additional mug~1 cm6020
Additional beverage bottle~2 cm5020

These results extend the evidence to three OOD objects. We will include these evaluations in Appendix E.6.

  1. Loss ablation. We trained a parameter-only Smooth-L1 head and a head using the composite objective in Eq. (32), each for 100 epochs. Each evaluated over 320 episodes spanning all tasks.
Training ObjectiveIoU ↑Boundary Gradient Cosine ↑TSR ↑CR ↓SSR ↑ETS ↓
Parameter-only regression0.74460.935857.1911.8855.00281.04
Composite objective, Eq. (32)0.76270.939160.9410.6259.38273.43

The composite objective improves SSR by 4.38 pp and reduces ETS by 7.61 steps, alongside lower CR and improved geometric metrics. This supports the complete objective, including boundary weighting and barrier supervision. We will add the results to the revision.

AuthorsOpenReview ↗

Response-3

Q2: Multiple Obstacles and Larger Held Objects

A: The current implementation predicts one primary obstacle and is aligned with the SafeLIBERO protocol. However, our control interface supports multi-obstacle expansion.

Perception Extension: Extending the geometry predictor to M obstacles only requires expanding the MLP output layer to regress K candidate ellipsoid slots (trained with set-based matching on multi-obstacle data), adding negligible forward-pass latency.

Control Scalability: Mathematically, our unified CBF-QP formulation (Eq. 9 & Appendix B.1 Eqs. 20–22) scales linearly with the number of active proxy-obstacle pairs. Specifically, for M obstacles, the proximity threshold hi,m,t<hfree(σt)h_{i,m,t} < h_{\text{free}}(\sigma_t) filters out distant obstacles, and each remaining active pair (i,m)∈Kt(i, m) \in K_t simply appends one auxiliary direction variable uz,i,m,t∈R3u_{z,i,m,t} \in \mathbb{R}^3 to the augmented decision vector and one linear inequality row Ai,mcbfut≥bi,mcbfA_{i,m}^{\text{cbf}}u_t \ge b_{i,m}^{\text{cbf}} to the QP.

For held objects, the current proxy is verified to enclose the evaluated objects. A larger but graspable object may extend beyond this envelope. In that case, separation of the proxy does not establish clearance for the uncovered geometry. Supporting such objects requires an object-dependent envelope that accounts for size, orientation, and offset relative to the gripper. We will clarify this operating boundary and present adaptive held-object geometry as future work.

Q3: Control Timing, Geometry Age, and Physical Training Data

A: We are delighted to clarify the question for you.

  1. Control Timing. As detailed in Appendix C.4 and E.2 (Tables 6 & 7, Fig. 10), the control interval is Δt=50 ms\Delta t = 50\text{ ms} (20 Hz). The local safety layer executes synchronously at every control step in 10.21 ms (Phase tracking: 0.03 ms, Proxy updates: 0.34 ms, Constraint construction: 2.79 ms, QP solve: 6.92 ms, Post-processing: 0.13 ms). Including remaining control overhead, the total mean step latency is 46.90 ms (<50 ms< 50\text{ ms} deadline), sustaining a 21.4 Hz average control rate.

  2. Geometry Estimate Staleness. As noted in Appendix E.2, 323 geometry queries occur across 1,600 control steps. Thus, visual obstacle predictions (ct,rt)(c_t, r_t) update once per action chunk (every ~5 steps or ≈250 ms\approx 250\text{ ms}) at an amortized latency of 12.63 ms/step. Crucially, because tabletop obstacles are static during a rollout and the robot's proxy poses (pi,t,Ri,t)(p_{i,t}, R_{i,t}), barrier values hi,th_{i,t}, and gradients h˙i,t\dot{h}_{i,t} are recomputed synchronously every 50 ms using live proprioception, this 1–5 step visual interval introduces no safety lag.

  3. Physical geometry supervision. Collecting physical geometry supervision is lightweight and requires no frame-by-frame manual annotation. After establishing the tabletop-to-base transformation, we measure an obstacle's base-frame bounding box. For box half-extents bkb_k, an enclosing ellipsoid with the same center and axes is constructed using rk⋆=3,bkr_k^\star=\sqrt{3}\\,b_k. This construction encloses the measured box. A scene label is reused while the obstacle pose remains unchanged. Pure translation requires updating the center; rotation additionally requires updating the bounding geometry. | Item | Physical dataset | |:--|:--| | Training configurations | 25 static configurations, 5 per task template, separate from evaluation | | Supervised samples | ~ 15,000 control-step samples | | Training objects | Bowl, Cola, cup, kettle, and milk | | Unseen evaluation instances | Wine bottle and the additional mug and beverage bottle reported in Q1 | | Manual measurement and annotation | ~ 5 hours |

We will add these dataset and annotation details to Appendix E.6 and clarify the control and geometry-update settings in Appendix C.4.


To sum up:

We clarified GeoPhase’s technical contributions beyond the established barrier primitives (W1). We provided outcome analyses and low-risk evaluations to quantify the aggregate safety–performance trade-off (W2). We added a jointly tuned static baseline to assess the benefit of phase-aware scheduling (W3). We supplemented the physical evaluation with AEGIS, demonstrating lower CR and higher SSR across all tasks (W4). We expanded unseen-object evaluation and compared the composite loss with geometry regression (Q1). We outlined the multi-obstacle extension and clarified the requirements for larger held objects (Q2). We detailed the reported runtime measurements, geometry-update schedule, physical training data, and annotation procedure (Q3).

We sincerely thank the reviewer for the constructive feedback, which has strengthened our empirical evaluation and clarified the method’s scope.

If you have any further questions or additional concerns, please feel free to discuss them with us.

Dear Reviewer JY75,

We truly value all of your suggestions. If you have any further questions, please do not hesitate to let us know. We will do our best to address them during the remainder of the discussion period. Thank you again for your time and effort. Best of luck!

Best regards, Authors

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

GeoPhase is a runtime safety filter for frozen VLA policies. It combines learned obstacle geometry, phase-dependent constraint scheduling, and multiple robot–object proxies within a CBF-QP that adjusts Cartesian actions. Experiments demonstrate improved collision avoidance and safe task completion in simulation and on a physical robot.

Critical strengths
  1. This method covers the held object and arm links as well as the end-effector, while keeping the VLA policy frozen. This addresses cases where the gripper clears an obstacle but the carried object or another arm link still collides with it. The proxy ablation supports this design: adding held-object and arm protection reduces CR from 29.4% to 12.8%.

  2. GeoPhase has the lowest average CR (12.8%) and highest average SSR (62.6%) among the reported methods on SafeLIBERO. Results on five tasks with a real robot also show improvements over the unfiltered policy: CR drops from 64.0% to 12.0%, while SSR increases from 18.0% to 74.0%.

Critical weaknesses
  1. It is not clear what is new in the constraint design. The pairwise barrier (Eq. 8) has the same form as the one in AEGIS [1]; both are based on the rotating supporting-hyperplane CBF of Funada et al. [2]. AEGIS [1] and KNOWS [3] also adapt the separating direction online, and KNOWS [3] already uses VLA attention to decide which objects are obstacles. The new parts seem to be the extra proxies and the phase schedule, and the paper should explain more clearly what they add.

  2. The phase ablation is not a fair comparison. The phase-aware parameters were tuned (Appendix E.4), but the static baseline simply uses α = 15 in every phase, and its thresholds are not given. I could not find any tuning for it. Without a tuned static baseline, it is hard to tell whether the improvement comes from scheduling or from tuning.

  3. I am not sure collision rates are measured the same way for all methods. The SafeDojo [4] and KNOWS [3] numbers are taken from their papers, and both papers describe collisions in terms of contact; KNOWS counts contact with any non-target obstacle. The collision check in the SafeLIBERO repository [5] only tracks whether one designated obstacle has moved. The paper does not say which check it uses, and the matching π0.5 numbers in Appendix D.3 only cover that one baseline. Please state the detector, the thresholds, and which objects are checked.

Questions for authors
  1. Technically, what do the extra proxies and the phase schedule add over AEGIS [1] and KNOWS [3]? Could you report a static baseline whose gain and thresholds are tuned on the same validation set as the phase-aware version?

  2. In simulation, does your collision detector check contact or obstacle displacement? Is it the same for GeoPhase, the reproduced AEGIS [1], and the SafeDojo [4] and KNOWS [3] numbers?

References

[1] S. Hu et al. “VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer.” arXiv:2512.11891, 2025. https://arxiv.org/abs/2512.11891

[2] R. Funada, K. Nishimoto, T. Ibuki, and M. Sampei. “Collision Avoidance for Ellipsoidal Rigid Bodies With Control Barrier Functions Designed From Rotating Supporting Hyperplanes.” IEEE Transactions on Control Systems Technology, 33(1):148–164, 2025. https://doi.org/10.1109/TCST.2024.3467809

[3] S. Park, F. Zhang, B. Mirzasoleiman, S. Talebi, and N. Sehatbakhsh. “Your Model Already Knows: Attention-Guided Safety Filter for Vision-Language-Action Models.” arXiv:2606.09749, 2026. https://arxiv.org/abs/2606.09749

[4] K. Tang et al. “SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model.” arXiv:2606.20698, 2026. https://arxiv.org/abs/2606.20698v1

[5] THURCSCT. SafeLIBERO dataset README. Hugging Face. https://huggingface.co/datasets/THURCSCT/SafeLIBERO/blob/main/README.md (accessed October 2026).

3 comments
AuthorsOpenReview ↗

Response-1

We sincerely thank the reviewer JXHV for the careful reading and constructive assessment. We are encouraged that you recognize the importance of improving geometric safety for VLA manipulation, and we provide detailed responses to each of your comments below.

W1 & Q1(a): Technical novelty and advances over AEGIS/KNOWS

A: We agree that Eq. (8) inherits the rotating supporting-hyperplane barrier from Funada et al. However, GeoPhase does not claim a new pairwise barrier primitive. Its contribution is a phase-aware, multi-proxy safety formulation that addresses limitations not jointly handled by these methods.

  1. Multi-proxy kinematic coupling under a shared Cartesian action.

The reported formulations of AEGIS and KNOWS focus on EEF–obstacle separation. In contrast, GeoPhase jointly represents the EEF, the held object, and body links. Because the frozen VLA outputs a Cartesian EEF command rather than independent link-level controls, these proxies must be coupled explicitly. Appendix B.1 (Eqs. 13–22) derives the link-to-action map Tℓ(i)←e=Jℓ(i)Je†T_{\ell(i)\leftarrow e}=J_{\ell(i)}J_e^\dagger together with the rigid-offset coupling matrix SiS_i. Under the stated local minimum-norm kinematic approximation, the spatial motion of every active proxy and its pair-specific separating-direction update {uz,i,t}i∈Kt\{u_{z,i,t}\}_{i\in\mathcal{K}_t} becomes part of a single affine constraint system over an augmented decision variable. Thus, one QP can coordinate the shared Cartesian action across the EEF, held-object, and body proxies.

Table 4 and Fig. 9 show that adding held-object and body proxies reduces CR from 29.4% (EEF-only) to 21.9% and 12.8%, respectively, while EEF intervention remains nearly unchanged (71.6–72.9%). Their intervention rates of 38.7% and 24.0% suggest protection against collision modes beyond those captured by the EEF proxy.

  1. Phase-aware constraint scheduling.

Different manipulation phases impose different geometric and task-level requirements: approach and grasp require sufficient proximity to the target, transport introduces the held object into the protected geometry, and place requires precise release. GeoPhase therefore schedules the active proxy set PσtP_{\sigma_t}, the activation threshold hfree(σt)h_{\mathrm{free}}(\sigma_t), and the barrier gain αi(σt)\alpha_i(\sigma_t) using an online finite-state machine.

The validation-tuned static baseline achieves SSR/CR of 59.6%/15.1% (See W2&Q1), compared with 62.6%/12.8% for phase-aware scheduling, demonstrating gains beyond fixed-parameter tuning. At the grasp-to-transport transition, GeoPhase additionally verifies the barrier initialization condition before activating the held-object proxy.

  1. Direct obstacle-geometry prediction from VLA embeddings.

KNOWS uses VLA attention to identify the policy’s intended target and still relies on external tracking and explicit 3D perception for obstacle geometry. GeoPhase directly regresses the obstacle ellipsoid parameters (ct,rt)=fθ(Zt)(c_t,r_t)=f_\theta(Z_t) from the pooled intermediate visual embedding ZtZ_t of the frozen VLA. The subsequent geometric constraints are computed analytically, without an explicit point-cloud reconstruction and ellipsoid-fitting pipeline.

We will revise the manuscript to distinguish these inherited components from our contributions.

AuthorsOpenReview ↗

Response-2

W2 & Q1(b): Fairness of the phase-aware ablation and tuned static baseline

A: We thank the reviewer for identifying the missing static-baseline tuning experiment. The original ablation used α=15s−1\alpha=15 s^{-1} and hfree=0.10mh_{\mathrm{free}}=0.10 m without reporting a dedicated search over static parameters.

To address this, we evaluated a global static baseline on the same validation set and with a matched tuning protocol. We searched αstatic∈�0�s−1,hfree,static∈�1�m\alpha_{\mathrm{static}}\in�0�s^{-1},h_{\mathrm{free,static}}\in�1�m:

αstatic\alpha_{\mathrm{static}} / hfree,statich_{\mathrm{free,static}}0.030.060.10
2.554.7 / 21.656.1 / 20.357.0 / 19.5
557.6 / 19.459.2 / 17.860.1 / 16.4
1058.8 / 17.460.9 / 15.561.9 / 14.1
1557.9 / 18.260.0 / 16.261.4 / 14.3
2055.6 / 20.158.3 / 18.059.5 / 16.5

We then evaluated the validation-optimal configuration (αstatic=10 s−1,hfree, static=0.10 m\alpha_{\text{static}} = 10\text{ s}^{-1}, h_{\text{free, static}} = 0.10\text{ m}) on the full 1,600-episode SafeLIBERO test set:

Methodα\alpha (s−1\mathrm{s}^{-1})hfreeh_{\mathrm{free}} (m)SSR ↑\uparrow (%)CR ↓\downarrow (%)
Original static baseline150.1059.416.2
Tuned global static100.1059.615.1
Phase-aware GeoPhase(10,10,15,5)(10,10,15,5)(0.10,0.06,0.10,0.03)(0.10,0.06,0.10,0.03)62.612.8

This confirms that the improvement arises from phase-aware scheduling rather than tuning bias: a single static configuration cannot simultaneously accommodate the permissive proximity required in Grasp/Place and the expanded protective envelope needed during Transport.

We will incorporate these results into Section 5.3 and Appendix E.4.

W3 & Q2: Collision-detector Specification

A: We appreciate the reviewer’s careful scrutiny of the collision metric. We agree that the detector and its scope should be stated explicitly.

  1. Detector, Threshold, and Monitored Objects.

Our evaluation follows the published collision check used by the SafeLIBERO benchmark. For each scene, the benchmark designates one safety-critical non-target obstacle. Let p0p_{0} be its initial position and ptp_t its position at control step tt. We flag a collision if, at any monitored step, d_t = \left\\|p_t - p_0\right\\|_1 = \sum_{j \in \\{x,y,z\\}} \left|p_{j,t} - p_{j,0}\right| > 0.001\\,\mathrm{m}

The same detector, threshold, initialization, and monitored-object selection are used for the locally evaluated unfiltered π0.5\pi_{0.5}, AEGIS, and GeoPhase.

  1. About SafeDojo and KNOWS.

SafeDojo states that its safety supervision is obtained from “SafeLIBERO contact-collision auditing” (Appendix J, Safety labels and training). KNOWS defines collision as whether “the episode contacted any non-target obstacle” (Section 4.1, Metrics). While described as "contact" in the text, we infer that it aligns with the SafeLIBERO 1 mm1\text{ mm} displacement check.

Because the official implementations of SafeDojo and KNOWS were unavailable, we used their published results.

We will revise Appendices D.3–D.4 to specify the detector implementation, threshold, and alignment verification.


To sum up:

In the discussion above, we clarified that GeoPhase’s novelty lies in phase-aware multi-proxy kinematic coupling and direct obstacle-geometry prediction from VLA embeddings (W1/Q1(a)). We provided a validation-tuned global static baseline and showed that phase-aware scheduling improves the safety–task-completion trade-off beyond parameter tuning (W2/Q1(b)). We also specified the local collision detector (W3/Q2). We sincerely thank the reviewer for the constructive feedback, which has helped us improve the precision and completeness of our presentation.

If you have any further questions or additional concerns, please feel free to discuss them with us.

Dear Reviewer JXHV,

We truly value all of your suggestions. If you have any further questions, please do not hesitate to let us know. We will do our best to address them during the remainder of the discussion period. Thank you again for your time and effort. Best of luck!

Best regards, Authors

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

Related papers

Discussion (0)

Sign in to comment.