acceptodds
Under review as a conference paper at ICLR 2027

Where to Stand for Manipulation: Affordance Prediction and Feasibility-Guided Standing Refinement for Last-Meter Mobile Manipulation

Abstract

Last-meter mobile manipulation requires a robot not only to navigate toward a task-relevant object, but also to terminate at a standing pose that enables successful downstream manipulation. However, existing navigation objectives primarily guide the robot to the target vicinity, without explicitly accounting for whether the resulting standing pose is feasible for the subsequent manipulation. We propose StandReady, which predicts manipulation-ready standing affordances from a single egocentric observation and a task instruction. To incorporate downstream manipulation feasibility, StandReady introduces manipulation-feasibility-guided standing refinement, which evaluates candidate standing poses using reference end-effector trajectories during training and uses the resulting trajectory-level rewards to refine affordance supervision. Experiments on the RoboCasa benchmark demonstrate that StandReady selects more manipulation-ready terminal poses and consistently improves downstream manipulation success over existing baselines. Ablation studies confirm that both feasibility-guided refinement and the trajectory-level reward design contribute to the performance gains. Overall, StandReady bridges the gap between navigation-oriented terminal-pose selection and the requirements of downstream manipulation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Mobile Manipulation, Robot Learning, Affordance Learning, Last-Meter Navigation, Robot Manipulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NSZ4OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes StandReady, a framework for predicting manipulation-ready terminal standing poses in last-meter mobile manipulation. Given a single egocentric RGB-D observation and a task instruction, StandReady predicts floor-level standing affordances together with the corresponding mobile-base orientations.

Critical strengths
  • The paper addresses a meaningful and practically relevant problem. A navigation-valid pose does not necessarily provide a suitable initial condition for downstream arm execution. Explicitly modeling the relationship between terminal base placement and subsequent manipulation is important for reliable mobile manipulation.

  • The proposed training-time refinement is well motivated and practically appealing. Compared with direct floor-level affordance prediction such as NavAff, StandReady incorporates trajectory-level, embodiment-aware feasibility signals, including reachability, IK feasibility, manipulability, Jacobian conditioning, and base compatibility. These signals are used during training while preserving a simple single-forward inference procedure at deployment.

  • The empirical results provide reasonable support for the proposed idea. StandReady improves the average RoboCasa success rate over NavAff and shows gains across the five evaluated tasks. The reward-component ablations, controlled arm-tracking study, and real-world terminal-pose evaluation further support the claim that unfavorable arm-motion conditions can affect downstream execution. The comparison with Mobi-π\pi also demonstrates a clear test-time efficiency advantage.

Critical weaknesses
  • The evaluation scope is still relatively limited. The main experiments focus on five atomic tasks in RoboCasa and use a fixed downstream policy. The paper does not evaluate generalization to unseen tasks, objects, environments, robot embodiments, or substantially different manipulation policies. The real-world study measures whether the terminal pose appears manipulation-ready rather than whether the robot completes the actual drawer-closing or door-closing task. Thus, the current results support the effectiveness of StandReady in the evaluated setting, while broader generality remains less established.

  • The method contains several heuristic design choices that are not fully analyzed. The refinement procedure samples candidates primarily from regions already assigned high affordance scores, which may limit its ability to recover feasible poses missed by the initial VQA prior. The paper does not provide sensitivity studies for the candidate budget, search radius, affordance threshold, reward weights, refinement coefficient, or number of refinement rounds. Stronger comparisons with an oracle feasibility search, a direct analytical IK-based baseline, or direct distillation from the best feasible candidate would also help isolate the contribution of the proposed iterative refinement procedure.

Questions for authors

See weakness.

7 comments
AuthorsOpenReview ↗

Response to Reviewer NSZ4[Part 1]

Dear Reviewer NSZ4,

Thank you for your constructive, thorough, and encouraging review. We are delighted that you found our paper clear and easy to follow, recognized the practical value of incorporating trajectory-level feasibility without test-time optimization, and appreciated the motivation behind bridging the gap between navigation and downstream manipulation.

Your feedback provided the primary roadmap for substantially strengthening our revision. Specifically, we have:

  1. Broadened our evaluation scope from 5 atomic tasks to 31 simulation tasks across two benchmarks (RoboCasa365 and ManiSkill3), three robotic arms (Franka Panda, Kinova3, UR5e), two VLA policies (π0.5\pi_{0.5} and XR-1), and six real-world household tasks evaluated via physical policy execution.
  2. Conducted extensive diagnostic studies and sensitivity analyses, including hyperparameter sweeps over search radius, candidate budget, and reward weights (Appendix E, Table 6), comparisons against analytical IK filtering and demonstration bounds (Section 4.2, Table 1), and an in-depth comparative analysis between iterative refinement and direct distillation from the best feasible candidate.

For a fully rendered version comparing our revisions against the initial submission, please refer to Revision_Summary.pdf in the Supplementary Material. In this document, newly added or substantially revised text is highlighted in blue, removed text is indicated in red strikethrough, and updated tables and figures are marked with blue revision tags in their captions.

Below, we address each weakness in detail.


W1: Evaluation scope and generalization across unseen tasks, embodiments, policies, and real-world task completion. "The evaluation scope is still relatively limited. The main experiments focus on five atomic tasks in RoboCasa and use a fixed downstream policy. The paper does not evaluate generalization to unseen tasks, objects, environments, robot embodiments, or substantially different manipulation policies. The real-world study measures whether the terminal pose appears manipulation-ready rather than whether the robot completes the actual drawer-closing or door-closing task. Thus, the current results support the effectiveness of StandReady in the evaluated setting, while broader generality remains less established."

Response: You raise a critical observation regarding the generality of our framework. We fully agree that evaluating only five atomic tasks on a single arm with static readiness checks leaves open questions about broader applicability. A robust mobile manipulation system must generalize across novel task instructions, diverse mechanical embodiments, different downstream policies, and messy physical environments.

In our original submission, we focused on five atomic tasks to validate the core mechanism of feasibility refinement in a controlled environment. Prompted by your critique, we recognized the necessity of establishing broader generality. In the revised manuscript, we substantially expanded our experimental suite across four operational axes:

  1. Task and environment breadth (31 simulation tasks across 2 simulators): We scaled from 5 to 25 tasks in RoboCasa365 (15 seen and 10 novel unseen tasks across 5 kitchen scenes) and added 6 novel pick-and-place tasks in ManiSkill3. The tasks cover diverse physical mechanics: hinged opening/closing, prismatic sliding, rotational fixture controls, and multi-stage pick-and-place transfers (with the complete task list detailed in Table 5 of the Appendix).
  2. Cross-embodiment generalization (3 robot arms): We benchmarked StandReady on Franka Emika Panda (7-DoF), Kinova3 (7-DoF), and UR5e (6-DoF), showing that the model discovers customized standing poses adapted to the distinct reach envelopes and joint geometries of each arm (Table 2 of the revised manuscript).
  3. Cross-policy transfer (2 VLA backbones): We evaluated downstream manipulation under both π0.5\pi_{0.5} and XR-1 policies, confirming that the spatial advantages of feasibility-guided standing hold across different policy architectures (Table 2 of the revised manuscript).
  4. Physical policy execution (6 real-world household tasks): We completely replaced our preliminary readiness study with physical policy execution on a four-wheel-steering mobile manipulator with an AgileX Piper arm under the XR-1 policy. Across 54 physical trials spanning three articulated fixtures and three pick-and-place tasks, StandReady achieves an overall task completion rate of 59.3%, exceeding N2M at 53.7% by 5.6 percentage points and NavAff at 35.2% by 24.1 percentage points (Table 4 of the revised manuscript). Representative video recordings of the real-world manipulation experiments are included in the Supplementary Material (realworld_lastmeter_moma_video.zip).

(Continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer NSZ4[Part 2]

The table below summarizes our expanded simulation benchmarking across seen and novel unseen task distributions:

Table R1: Task generalization across seen and novel unseen benchmarks under π0.5\pi_{0.5} and XR-1 policies (RoboCasa Seen pools 2,130 episodes across 15 tasks and 3 robot arms; RoboCasa Unseen pools 1,410 episodes across 10 novel tasks and 3 robot arms; ManiSkill3 evaluates 120 episodes across 6 novel tasks with Panda):

Benchmark SplitPolicyLeLaNNavAffGPT-6.1 SolN2MStandReady (Ours)VLA Baseline
RoboCasa Seen (15 tasks)π0.5\pi_{0.5}4.32%11.03%22.02%28.54%34.04%37.14%
XR-14.55%8.69%22.25%30.09%32.44%44.98%
RoboCasa Unseen (10 tasks)π0.5\pi_{0.5}9.08%11.84%21.49%29.65%41.77%46.17%
XR-110.99%12.41%18.87%29.86%42.48%49.22%
ManiSkill3 Unseen (6 tasks)π0.5\pi_{0.5}19.17%40.83%42.50%50.83%63.33%47.50%
XR-115.00%70.00%60.83%77.50%80.00%72.50%

As reported in Table R1, StandReady maintains consistent gains across all benchmark splits. The framework exhibits a particularly pronounced advantage on novel unseen tasks, where StandReady raises success over N2M from 29.65% to 41.77% under π0.5\pi_{0.5} (an improvement of 12.12 percentage points) and from 29.86% to 42.48% under XR-1 (an improvement of 12.62 percentage points). This confirms that the framework generalizes well beyond its initial training distribution.

Manuscript Revisions: Updated Section 1 (Paragraphs 5–6) and completely replaced the preliminary evaluation in Section 4.5 (Paragraphs 2–3) to reflect the expanded multi-task, multi-embodiment, and physical evaluation suite:

Section 1 (Introduction, Paragraph 5): "Finally, we evaluate StandReady in simulation and on a physical mobile manipulator. Simulation results demonstrate improved downstream manipulation success, and real-world experiments further validate physical transfer through downstream policy execution and task-completion evaluation across six household tasks."

Section 1 (Introduction, Paragraph 6, Item 3): "We evaluate StandReady on 25 RoboCasa365 tasks across three robot embodiments and two VLA policies, with additional cross-simulator experiments in ManiSkill3 and trials on a physical mobile manipulator, demonstrating improved downstream task success over representative baselines and validating the proposed refinement through ablation studies."

Section 4.5 (Real-World Experiments, Paragraphs 2–3): "We evaluate three articulated-object tasks and three pick-and-place tasks. For each task, each autonomous method predicts one terminal standing pose from a single observation at each of four or five initial approach locations. Each selected pose is evaluated in two trials, in which the robot navigates to the pose and executes the manipulation with XR-1."

"As shown in Table 4, StandReady achieves the highest overall success rate among the autonomous methods at 59.3% (32/54), followed by N2M at 53.7% (29/54) and NavAff at 35.2% (19/54). StandReady records one additional success over N2M on each pick-and-place task, accounting for all three additional successes overall. Human-selected poses yield 77.8% (42/54). Figure 6 shows that StandReady's predicted poses generally cluster in front of the target fixtures. N2M also produces clusters, while NavAff's poses are more dispersed. These results suggest that trajectory-feasibility priors learned in simulation can support standing-pose selection for real-world VLA execution."

(Continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer NSZ4[Part 3]

W2: Heuristic design choices, sensitivity analysis, and comparative baselines (analytical IK, oracle search, direct distillation). "The method contains several heuristic design choices that are not fully analyzed. The refinement procedure samples candidates primarily from regions already assigned high affordance scores, which may limit its ability to recover feasible poses missed by the initial VQA prior. The paper does not provide sensitivity studies for the candidate budget, search radius, affordance threshold, reward weights, refinement coefficient, or number of refinement rounds. Stronger comparisons with an oracle feasibility search, a direct analytical IK-based baseline, or direct distillation from the best feasible candidate would also help isolate the contribution of the proposed iterative refinement procedure."

Response: Your feedback highlights the need to systematically analyze the algorithmic mechanics of Stage-II refinement. We agree that relying on heuristic hyperparameters without sensitivity analysis leaves their stability uncertain, and that understanding the exact benefits of iterative refinement requires rigorous comparisons against analytical IK, oracle bounds, and direct distillation.

To address these concerns thoroughly, we structure our response into three complementary parts:

  1. Candidate sampling mechanics and hyperparameter sensitivity sweeps.
  2. Comparison against analytical IK filtering and oracle feasibility placement.
  3. In-depth analysis of iterative refinement compared with direct distillation on seen and unseen tasks.

Part 1: Candidate Sampling Mechanics and Sensitivity Sweeps

You raised a valid concern regarding whether sampling candidates from high-affordance regions might trap the search in suboptimal basins if the initial VQA prior misses the globally optimal standing area. In our framework, candidate sampling is explicitly designed to balance visual-semantic guidance with continuous kinematic exploration. Rather than collapsing onto a narrow peak, Stage I provides coarse semantic grounding that establishes a broad, permissive interaction zone across the floor in front of the target fixture while filtering out irrelevant open-scene areas. Within this task-relevant floor support, each refinement round allocates a budget of N=32N = 32 candidate standings across two complementary spatial scales: while 75% of the budget performs local exploitation within a near radius of 0.5 m around the incumbent prediction, 25% is strictly reserved for broader spatial exploration up to an outer search radius of 2.0 m (as formulated in Section 3.3 andAppendix D of the revised manuscript). This deliberate non-local allocation prevents the search from getting trapped in local extrema and allows Stage-II feasibility scoring to discover kinematically favorable floor locations even if they deviate from the initial VQA center.

To evaluate the operational stability of these design choices, we conducted controlled sweeps on a 5-task subset of RoboCasa Seen tasks across 250 evaluation rollouts, as detailed in Appendix E, Table 6 of the revised manuscript:

Table R2: Sensitivity analysis of Stage-II refinement hyperparameters across five categories (evaluated on the 5-task RoboCasa Seen Panda subset across 250 rollouts per setting). The default configuration uses candidate budget N=32N = 32, refinement rounds k=8k = 8, outer search radius 2.0 m (near radius 0.5 m), risk-to-dexterity weight ratio 0.7:0.3, and base-footprint penalty weight λb=1.0\lambda_b = 1.0, achieving Avg. Sfea=0.5942S_{\mathrm{fea}} = 0.5942 and 46.0% policy success. Each block varies one factor from this default baseline. Relative Step Time measures the training time per refinement step normalized to the default configuration (100%). In the table, dashes under Success Rate indicate settings where rollouts were omitted to conserve compute, while dashes under Avg. SfeaS_{\mathrm{fea}} indicate altered reward weights whose raw values lie on a different scale:

(Continued in the next comment)

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

This paper proposes StandReady, a framework for predicting manipulation-ready standing poses for last-meter mobile manipulation. It first learns standing affordances from VLM-generated supervision and then refines them using trajectory-level manipulation feasibility, including arm reachability, manipulability, and base collision constraints. Experiments on RoboCasa demonstrate improved downstream manipulation success over several baselines, while limited real-world experiments evaluate terminal standing readiness.

Critical strengths
  1. Important problem with a practical solution. The paper addresses a meaningful challenge in mobile manipulation: selecting a terminal standing pose that facilitates downstream manipulation rather than merely reaching the target vicinity. The proposed offline refinement allows efficient single-forward inference at deployment.

  2. Reasonable simulation results. The method improves downstream manipulation success on five RoboCasa tasks, and the ablation studies support the contribution of feasibility-guided refinement.

Critical weaknesses
  1. Insufficient validation of the trajectory-level feasibility metric. The proposed feasibility reward primarily considers kinematic factors such as reachability, manipulability, and Jacobian conditioning. However, successful manipulation also depends on visual observations, contact dynamics, and the characteristics of the downstream policy. The paper does not directly establish how well the feasibility score predicts actual manipulation success across different candidate standing poses. Therefore, the validity of this surrogate for downstream policy performance remains insufficiently demonstrated.

  2. Limited real-world validation. Although the paper claims to improve manipulation-ready standing selection, its real-world evaluation only measures whether the selected poses provide plausible initial conditions for manipulation, without executing the downstream manipulation policy or reporting actual task success rates. Furthermore, the experiments cover only two tasks in relatively limited settings. This makes it difficult to assess whether the improvements observed in simulation translate into reliable real-world mobile manipulation.

Questions for authors

Q1. How strongly does the trajectory-level feasibility score correlate with actual downstream manipulation success across different standing poses? Could the authors provide an analysis comparing predicted feasibility scores with manipulation success rates obtained by executing the fixed downstream policy?

Q2. Why do the real-world experiments evaluate only standing readiness rather than actual manipulation success? Could the authors provide end-to-end real-world task success rates for StandReady and the baselines?

5 comments
AuthorsOpenReview ↗

Response to Reviewer qahy[Part 1]

Dear Reviewer qahy,

Thank you for your constructive and thorough review, and for recognizing the practical value of selecting manipulation-ready standing poses while retaining single-forward inference at deployment. Your feedback pointed directly to the two most critical evidentiary gaps in our original submission: verifying whether our kinematic trajectory feasibility score (SfeaS_{\mathrm{fea}}) genuinely predicts downstream manipulation success, and validating physical transfer through real manipulation execution rather than static standing-readiness checks.

To address these concerns directly, we carried out two major empirical additions in the revised manuscript:

  1. Large-Scale Validation of the Feasibility Metric: We evaluated our feasibility score against more than 27,000 rollouts under a fixed π0.5\pi_{0.5} policy across 15 tasks and 5 kitchen scenes (67 valid task–layout pairs) in RoboCasa, confirming a consistent monotonic correlation with task success even when holding target distance within narrow bands, supported by inverse-kinematics diagnostic comparisons.
  2. Real-World Task-Completion Evaluation: We replaced the preliminary two-task standing readiness evaluation with physical policy execution across six household manipulation tasks (54 trials per method) using the XR-1 policy, demonstrating higher task completion rates than existing baselines.

For a fully rendered version comparing our revisions against the initial submission, please refer to Revision_Summary.pdf in the Supplementary Material. In this document, newly added or substantially revised text is highlighted in blue, removed text is indicated in red strikethrough, and updated tables and figures are marked with blue revision tags in their captions.

Below, we address each weakness and question in detail.


W1: Insufficient validation of the trajectory-level feasibility metric. "The proposed feasibility reward primarily considers kinematic factors such as reachability, manipulability, and Jacobian conditioning. However, successful manipulation also depends on visual observations, contact dynamics, and the characteristics of the downstream policy. The paper does not directly establish how well the feasibility score predicts actual manipulation success across different candidate standing poses. Therefore, the validity of this surrogate for downstream policy performance remains insufficiently demonstrated."

Response: You raise an essential point. Because mobile manipulation involves perception, contact dynamics, and downstream policy behavior, an offline kinematic metric cannot be taken for granted without empirical grounding. As formulated in Section 3.3 (Eq. 4) of the revised manuscript: Sfea(a)=1−λmCmotion(a∣ξ)−λbCbase(a∣o)S_{\mathrm{fea}}(a) = 1 - \lambda_m C_{\mathrm{motion}}(a \mid \xi) - \lambda_b C_{\mathrm{base}}(a \mid o) This score is designed to evaluate candidate base placements by balancing waypoint-aggregated arm motion feasibility with base clearance. To determine whether it serves as a reliable surrogate for policy execution rather than an abstract kinematic index, we conducted an empirical validation across 15 manipulation tasks and 5 kitchen scenes (yielding 67 valid task–layout pairs with human demonstrations) in RoboCasa training environments, collecting over 27,000 rollouts under a fixed π0.5\pi_{0.5} manipulation policy.

First, we examined how downstream task success varies with SfeaS_{\mathrm{fea}} across candidate poses. Candidate standing poses were ranked by SfeaS_{\mathrm{fea}} within each task–layout pair and partitioned into five equal-sized score quintiles from lowest to highest, denoted as Q1 through Q5. As shown by the solid curve in Figure 3 of the revised manuscript, task success rises strictly monotonically with SfeaS_{\mathrm{fea}}. To confirm that this trend does not simply reflect proximity to the workstation, we stratified candidate poses into 0.10 m base-to-target distance bands. Even at comparable distances to the target, task success continues to increase monotonically across score quintiles, as indicated by the dashed curve, confirming that the score captures execution-relevant kinematic conditioning beyond distance alone.

Second, we evaluated how different kinematic criteria affect downstream policy rollouts under identical initial scene configurations. Across the same 67 task–layout pairs, the diagnostic comparison below evaluates candidate filtering strategies against explicit placement baselines:

(Continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer qahy[Part 2]

Manipulation success across standing-pose selection and filtering criteria (Franka Panda with π0.5\pi_{0.5} policy across 67 valid task–layout pairs from 15 tasks and 5 kitchen scenes in RoboCasa; top rows filter a random candidate pool of 33,500 rollouts within a target-front interaction window; bottom rows evaluate single standing placements chosen directly by each criterion with 10 trials per pair):

Base Placement ConditionSuccess / TotalSR (%)
Random (2.0×1.02.0 \times 1.0 m)8,826 / 33,50026.35
+ Initial-contact IK8,627 / 19,96543.21
+ Full-trajectory IK7,605 / 14,84751.22
Semantic geometry (10-view)364 / 67054.33
Demonstration base pose426 / 67063.58
Feasibility validator (Ours)427 / 67063.73

As shown above, filtering candidates across the entire reference motion substantially outperforms checking reachability only at the initial contact point, increasing the success rate from 43.21% to 51.22% (an improvement of 8.01 percentage points), confirming that endpoint feasibility alone fails to guarantee multi-step execution. Furthermore, directly maximizing SfeaS_{\mathrm{fea}} achieves 63.73% success, outperforming 10-view semantic geometry at 54.33% by 9.40 percentage points and closely matching the policy's demonstration origin at 63.58%. This parity confirms that our kinematic score reliably captures the spatial conditions under which the policy succeeds.

These results indicate that SfeaS_{\mathrm{fea}} does not merely select poses closer to the target; by penalizing ill-conditioned joint configurations and reachability bottlenecks along the trajectory, it identifies base placements that provide favorable starting conditions for downstream policy execution.

Manuscript Revisions: Added a dedicated diagnostic subsection (Section 4.2, Paragraphs 1–2), Figure 3, and Table 1 validating SfeaS_{\mathrm{fea}} against policy rollouts and placement baselines:

Section 4.2 (Feasibility Score Validation and Diagnostic Comparison, Paragraphs 1–2): "To evaluate whether Manipulation Feasibility Validation (Section 3.3) reliably reflects downstream manipulation conditions, we first examine the empirical correlation between its feasibility reward SfeaS_{\mathrm{fea}} (Eq. 4) and task success. Across 67 task–layout pairs, candidate standing poses are sampled within a predefined region, scored via SfeaS_{\mathrm{fea}}, and evaluated by executing the downstream policy. When candidate poses within each task–layout pair are ranked by SfeaS_{\mathrm{fea}} and partitioned into five equal-sized quintiles from lowest to highest, denoted as Q1 through Q5, task success increases monotonically with SfeaS_{\mathrm{fea}}, as shown by the solid curve in Fig. 3. Crucially, this upward trend persists when ranking poses within 0.10 m base-to-target distance bands, as indicated by the dashed curve, confirming that SfeaS_{\mathrm{fea}} captures genuine kinematic and geometric feasibility rather than merely serving as a proxy for target proximity."

"To isolate how different kinematic and geometric criteria affect downstream execution, Table 1 compares candidate filtering strategies and explicit placement baselines. In the top block, kinematic filtering substantially improves over random placement, which yields only 26.35% success. Furthermore, extending reachability checks from initial contact alone to the full reference trajectory increases the success rate from 43.21% to 51.22%, representing an improvement of 8.01 percentage points and confirming that single-waypoint reachability is insufficient. In the bottom block, maximizing SfeaS_{\mathrm{fea}} within 0.5 m of the demonstration origin achieves 63.73% success, outperforming 10-view semantic geometry at 54.33% by 9.40 percentage points and closely matching the policy's in-distribution demonstration pose at 63.58%. This parity confirms that SfeaS_{\mathrm{fea}} reliably captures execution-favorable standing conditions for supervising Stage II refinement."

(Continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer qahy[Part 3]

W2: Limited real-world validation. "Although the paper claims to improve manipulation-ready standing selection, its real-world evaluation only measures whether the selected poses provide plausible initial conditions for manipulation, without executing the downstream manipulation policy or reporting actual task success rates. Furthermore, the experiments cover only two tasks in relatively limited settings. This makes it difficult to assess whether the improvements observed in simulation translate into reliable real-world mobile manipulation."

Response: Your feedback prompted a valuable reflection on how we evaluate real-world mobile manipulation. In our initial design, we focused on terminal standing readiness to isolate the spatial handoff of the mobile base from the execution variations of the robotic arm. Your critique encouraged us to view this from a more holistic perspective: in physical deployment, a standing pose cannot be treated as an isolated geometric endpoint. Factors such as visual occlusion, physical contact dynamics, and arm compliance mean that the ultimate test of a standing pose is whether the downstream policy can actually complete the task from that pose.

Inspired by this perspective, we evolved our evaluation from static readiness checks to physical policy execution across six household manipulation tasks. We deployed our system on a four-wheel-steering mobile manipulator equipped with a 6-DoF AgileX Piper arm, dual head-mounted RGB-D cameras, a wrist camera, and LiDAR-inertial state estimation. Downstream manipulation was executed using the XR-1 policy. We compared StandReady against two representative single-forward standing-prediction baselines: NavAff (which predicts floor affordances using explicit object segmentation) and N2M (which predicts policy-preferred base poses from egocentric point clouds).

The evaluation spans six physical tasks: three articulated-fixture interactions (Close Drawer, Close Microwave, Close Cabinet) and three pick-and-place tasks (Pick-and-Place Mug to Sink, to Microwave, and to Drawer). For each task, the robot was initialized across four to five distinct approach locations. Each method predicted a terminal standing pose from a single egocentric observation, navigated to that pose, locked its mobile base, and executed the manipulation policy to completion over two repeated trials per pose (54 trials per method in total):

Real-world task-completion success rates under XR-1 policy execution on an AgileX Piper mobile manipulator (54 total trials per method across 6 household tasks; Humans serves as an upper reference using human-chosen base poses):

MethodClose DrawerClose Micro.Close CabinetP2P SinkP2P Micro.P2P DrawerOverall
NavAff20.0% (2/10)50.0% (4/8)0.0% (0/8)60.0% (6/10)62.5% (5/8)20.0% (2/10)35.2% (19/54)
N2M50.0% (5/10)62.5% (5/8)12.5% (1/8)80.0% (8/10)62.5% (5/8)50.0% (5/10)53.7% (29/54)
StandReady (Ours)50.0% (5/10)50.0% (4/8)25.0% (2/8)90.0% (9/10)75.0% (6/8)60.0% (6/10)59.3% (32/54)
Humans80.0% (8/10)75.0% (6/8)37.5% (3/8)90.0% (9/10)75.0% (6/8)100.0% (10/10)77.8% (42/54)

Across autonomous methods, StandReady achieves the highest overall success rate at 59.3% (32/54), with the clearest margin emerging on multi-stage pick-and-place tasks. Because trajectory-level feasibility accounts for both grasping and release waypoints, it prevents the base from stopping at poses where the arm can grasp the object but subsequently hits joint limits during placement. In addition, StandReady avoids the segmentation failures observed in NavAff on visually ambiguous fixtures such as Close Cabinet. Top-down pose overlays in Figure 6 of the revised manuscript show that StandReady’s predicted poses cluster consistently in front of the target interaction zones. For reference, human-selected standing poses achieve 77.8% success, reflecting the practical performance upper bound of the fixed XR-1 policy, where remaining failures stem from physical arm execution (such as gripper slip) rather than base placement. Representative video recordings of the real-world manipulation experiments are included in the Supplementary Material (realworld_lastmeter_moma_video.zip).

(Continued in the next comment)

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

This paper proposes StandReady, a task-conditioned standing pose prediction framework for last-meter mobile manipulation. The method consists of two stages: (1) learning standing affordances from VLM-generated supervision, and (2) refining the prediction model using manipulation feasibility scores based on reachability, manipulability, singularity, and base collision constraints. The framework enables efficient standing pose prediction without test-time optimization.

Critical strengths
  1. The paper is well-written, with a clear and easy-to-follow presentation of the proposed method.
  2. The inclusion of real-world robotic experiments helps demonstrate the practical applicability of the proposed approach.
  3. The proposed method predicts manipulation-feasible standing poses while reducing computational overhead by avoiding test-time optimization.
Critical weaknesses
  1. The paper uses VLM-generated annotations as supervision for Standing Affordance Prediction. However, it is unclear why the authors train a separate prediction model instead of directly querying the Qwen API at test time. Moreover, the quality of the learned standing affordances may be limited by the capabilities of the supervising VLM.
  2. The current ablation studies are insufficient to demonstrate the effectiveness of each component. In particular, the paper lacks an ablation without VQA-based supervision. It remains unclear what advantages the proposed framework offers over directly querying a lightweight VLM to predict standing poses.
  3. The paper does not sufficiently demonstrate the advantages of the proposed Stage II refinement over a simpler approach that combines VQA-based standing pose prediction with direct reachability checking of target manipulation points using cuRobo.
  4. The experimental comparisons do not include recent methods for standing pose selection, such as N2M.
Questions for authors
  1. The proposed method uses PointNet++ to encode point clouds. How robust is the model to noisy or incomplete point clouds captured by real-world depth sensors?
  2. In the qualitative comparisons in Fig. 4, why is the Turn On Faucet task evaluated in different scenes, while all other tasks are compared within the same scene?
  3. Does the Close Door task require whole-body motion planning with coordinated base and arm movements, rather than manipulation from a fixed base pose?
  4. Given that standing pose selection is a general problem in mobile manipulation, could the proposed model generalize to a broader range of mobile manipulation tasks?
7 comments
AuthorsOpenReview ↗

Response to Reviewer tYba[Part 1]

Dear Reviewer tYba,

Thank you for your constructive and thorough review, and for recognizing the clarity of our presentation, the practical efficiency of avoiding test-time optimization, and the value of our physical experiments. Your comments guided several key enhancements in our revision, including benchmarking a frontier multimodal foundation model (GPT-6.1 Sol) and N2M across our expanded evaluation suite, incorporating analytical inverse-kinematics diagnostics, clarifying our fixed-base problem formulation, and unifying the qualitative scenes in Figure 4.

For a fully rendered version comparing our revisions against the initial submission, please refer to Revision_Summary.pdf in the Supplementary Material. In this document, newly added or substantially revised text is highlighted in blue, removed text is indicated in red strikethrough, and updated tables and figures are marked with blue revision tags in their captions.

Below, we address each weakness and question in detail.

W1: Training a separate predictor vs. direct test-time VLM querying, and supervising VLM capability limits. "The paper uses VLM-generated annotations as supervision for Standing Affordance Prediction. However, it is unclear why the authors train a separate prediction model instead of directly querying the Qwen API at test time. Moreover, the quality of the learned standing affordances may be limited by the capabilities of the supervising VLM."

Response: Your feedback highlights an essential question regarding system architecture: why train a dedicated 3D affordance network rather than directly querying a VLM at test time, and how the model avoids being constrained by the supervising VLM's capability limits. We appreciate this observation, as it touches directly on the division of responsibilities between high-level foundation models and embodiment-specific physical execution.

In mobile manipulation, selecting an effective standing pose involves two distinct aspects: semantic task grounding (identifying the target fixture from language) and continuous 3D kinematic feasibility (ensuring the specific robot arm can complete downstream execution without collisions or singularity limits). While vision-language models excel at semantic association and 2D visual grounding, standard VLMs typically do not incorporate explicit, embodiment-specific kinematic models, such as continuous joint limits, analytical Jacobians, or link geometries. Querying a VLM online to optimize standing poses would require estimating 3D interaction waypoints from single-view observations and solving iterative inverse kinematics on the fly, which introduces substantial runtime latency and compounds visual perception errors. StandReady addresses this by decoupling the two roles: the supervising VLM provides coarse semantic grounding offline, while Stage-II kinematic refinement compiles embodiment-specific trajectory feasibility into a feedforward predictor, enabling single-forward inference (0.40 s) at deployment.

To evaluate direct test-time VLM querying empirically, we benchmarked GPT-6.1 Sol (OpenAI, 2026), a frontier multimodal foundation model, by prompting it zero-shot to predict task-conditioned standing floor pixels from egocentric observations across our benchmark splits:

Table R1: Manipulation success rate (%) of zero-shot VLM querying (GPT-6.1 Sol) versus N2M and StandReady (RoboCasa365 pools Panda, Kinova3, and UR5e across 2,130 episodes on Seen and 1,410 episodes on Unseen; ManiSkill3 evaluates Panda across 120 episodes):

MethodPolicyRoboCasa Seen (15 tasks)RoboCasa Unseen (10 tasks)ManiSkill3 Unseen (6 tasks)
GPT-6.1 Sol (Zero-shot VLM)π0.5\pi_{0.5}22.02%21.49%42.50%
XR-122.25%18.87%60.83%
N2M (Chai et al., ICML 2026)π0.5\pi_{0.5}28.54%29.65%50.83%
XR-130.09%29.86%77.50%
StandReady (Ours)π0.5\pi_{0.5}34.04%41.77%63.33%
XR-132.44%42.48%80.00%

(Continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer tYba[Part 2]

As reported in Table R1, two key findings emerge from the empirical comparison:

(1) Quantitative Comparison Across Benchmarks: Direct zero-shot VLM querying achieves only 22.02% on seen tasks and 21.49% on unseen tasks under π0.5\pi_{0.5}, trailing StandReady at 34.04% and 41.77% by 12.02 and 20.28 percentage points, respectively. Similar performance deficits appear under XR-1 and in ManiSkill3. Although GPT-6.1 Sol provides plausible semantic grounding near target fixtures, it lacks the 3D kinematic reasoning required to optimize base distance and heading for multi-step arm execution.

(2) Functional Role of Stage-I VQA Supervision vs. Stage-II Refinement: Stage-I VQA supervision is not intended to provide millimeter-accurate placement coordinates, but rather coarse floor-plane priors. Crucially, Stage-I VQA annotations were collected on only 5 atomic tasks. Among the 15 seen evaluation tasks, only 3 overlap with this initial set, while the remaining 12 seen tasks and all 10 novel unseen tasks received zero task-specific VQA supervision. StandReady transfers these semantic priors zero-shot to unfamiliar tasks, while Stage-II refinement independently optimizes physical feasibility using robot kinematic models and reference trajectories.

Manuscript Revisions: Updated Section 4.3 (Paragraph 2) of the revised manuscript to contrast direct foundation model querying with dedicated kinematic standing prediction:

Section 4.3 (Main Results, Paragraph 2): "Analyzing comparative methods highlights clear architectural trade-offs across standing-selection paradigms. Given only a single egocentric observation, frontier foundation models like GPT-6.1 Sol provide reasonable semantic grounding, achieving an overall success rate of 22.02% on seen tasks and 21.49% on unseen tasks under π0.5\pi_{0.5}, but lack the 3D kinematic awareness necessary to reason about joint limits, arm reachability, or base collisions. Meanwhile, although N2M benefits from dedicated per-task training on seen tasks, its lack of instruction conditioning necessitates borrowing proxy weights on novel unseen tasks, leading to policy-pose misalignment. Conversely, search-based methods face both representational and computational bottlenecks... In contrast, feedforward prediction models like StandReady and N2M eliminate test-time optimization entirely, requiring an average execution time of only 0.40 s and 0.52 s per episode, respectively..."


W2: Ablation without VQA supervision and advantages over querying a lightweight VLM. "The current ablation studies are insufficient to demonstrate the effectiveness of each component. In particular, the paper lacks an ablation without VQA-based supervision. It remains unclear what advantages the proposed framework offers over directly querying a lightweight VLM to predict standing poses."

Response: We agree that our initial submission lacked an ablation that completely removes Stage-I VQA supervision. We appreciate this observation, as it prompted us to systematically examine both the structural necessity of Stage-I grounding and the distinction between our model and lightweight 2D VLMs across two complementary aspects:

(1) Structural Role of Stage-I Semantic Grounding: In our two-stage architecture, Stage I provides coarse spatial grounding over the observed floor plane. Completely removing this prior would leave candidate sampling ungrounded across the open 3D scene, forcing Stage-II kinematic checks to evaluate thousands of physically irrelevant floor locations. To evaluate whether our framework's performance merely inherits the supervising VLM's capabilities, our benchmark against GPT-6.1 Sol in Table R1 demonstrates that pure VLM querying remains insufficient without 3D kinematic refinement. This confirms that the performance bottleneck lies in the lack of explicit 3D kinematic reasoning rather than the capacity of the supervising VLM, even when querying a frontier multimodal foundation model such as GPT-6.1 Sol.

(2) Distinct Kinematic Advantages Over Lightweight 2D VLMs: Lightweight or edge-deployed VLMs share a similar operational boundary with general foundation models: their representations are primarily optimized for pixel-level visual-semantic reasoning rather than evaluating continuous joint limits, manipulability, or physical footprint collisions for a specific embodiment. By fusing language-conditioned visual queries with 3D point-cloud geometry via PointNet++ and refining affordances against embodiment-specific trajectories, StandReady discovers standing poses customized to distinct arm kinematics. For instance, on the Kinova3 arm under π0.5\pi_{0.5}, StandReady improves task success from 27.61% to 36.90% on seen tasks and from 37.87% to 44.04% on unseen tasks, exceeding the Panda-calibrated demonstration baseline by 9.29 and 6.17 percentage points, respectively (as shown in Table 2 of the revised manuscript).

(Continued in the next comment)

AuthorsOpenReview ↗

Response to Reviewer tYba[Part 3]

W3: Advantages over combining VQA standing prediction with direct reachability checking via cuRobo. "The paper does not sufficiently demonstrate the advantages of the proposed Stage II refinement over a simpler approach that combines VQA-based standing pose prediction with direct reachability checking of target manipulation points using cuRobo."

Response: You raise an insightful baseline suggestion. Combining 2D VLM standing proposals with direct test-time reachability checking (such as via cuRobo) represents an intuitive motion-aware strategy, which we discuss in Related Work (Section 2 of the revised manuscript). In our initial formulation, we sought to avoid test-time kinematic optimization entirely by compiling feasibility priors into an offline-trained feedforward predictor.

Your comment encouraged us to examine the exact performance difference between checking initial target-point reachability versus full-trajectory kinematic feasibility. Operationally, test-time reachability checking requires accurate 3D target coordinates, whereas in last-meter navigation the robot receives only a single egocentric RGB-D observation and an instruction. Estimating waypoints from noisy single-view point clouds introduces compounding perception errors, and cuRobo optimization incurs online latency. StandReady avoids this by transferring feasibility knowledge offline during training, maintaining feedforward single-pass inference (0.40 s) at test time.

Furthermore, checking reachability only at the initial contact point fails to capture whether the arm can complete subsequent multi-step manipulation motions without hitting joint limits or singular configurations. To isolate this effect directly, our diagnostic study evaluated candidate base placements across 67 task–layout pairs in RoboCasa training scenes:

Table R2: Manipulation success across standing-pose selection and filtering criteria (Franka Panda with π0.5\pi_{0.5} policy across 67 valid task–layout pairs from 15 tasks and 5 kitchen scenes in RoboCasa; top rows filter a random candidate pool of 33,500 rollouts within a target-front interaction window; bottom rows evaluate single standing placements chosen directly by each criterion with 10 trials per pair):

Base Placement ConditionSuccess / TotalSR (%)
Random (2.0×1.02.0 \times 1.0 m)8,826 / 33,50026.35
+ Initial-contact IK8,627 / 19,96543.21
+ Full-trajectory IK7,605 / 14,84751.22
Semantic geometry (10-view)364 / 67054.33
Demonstration base pose426 / 67063.58
Feasibility validator (Ours)427 / 67063.73

As reported in Table R2, extending reachability checks from the initial contact point (43.21%) across the full reference trajectory (51.22%) increases the success rate by 8.01 percentage points. This confirms that single-waypoint reachability overlooks mid-trajectory singularities and workspace boundaries. Furthermore, directly maximizing SfeaS_{\mathrm{fea}} achieves 63.73% success, outperforming 10-view semantic geometry at 54.33% by 9.40 percentage points and closely matching the policy's demonstration origin at 63.58%. This parity confirms that our kinematic score reliably captures execution-favorable standing conditions without requiring test-time optimization.

Manuscript Revisions: Added a dedicated diagnostic subsection (Section 4.2), Figure 3, and Table 1 in the revised manuscript:

Section 4.2 (Feasibility Score Validation and Diagnostic Comparison, Paragraph 2): "To isolate how different kinematic and geometric criteria affect downstream execution, Table 1 compares candidate filtering strategies and explicit placement baselines. In the top block, kinematic filtering substantially improves over random placement, which yields only 26.35% success. Furthermore, extending reachability checks from initial contact alone to the full reference trajectory increases the success rate from 43.21% to 51.22%, representing an improvement of 8.01 percentage points and confirming that single-waypoint reachability is insufficient. In the bottom block, maximizing SfeaS_{\mathrm{fea}} within 0.5 m of the demonstration origin achieves 63.73% success, outperforming 10-view semantic geometry at 54.33% by 9.40 percentage points and closely matching the policy's in-distribution demonstration pose at 63.58%. This parity confirms that SfeaS_{\mathrm{fea}} reliably captures execution-favorable standing conditions for supervising Stage II refinement."

(Continued in the next comment)

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their thoughtful and constructive feedback. Your comments have helped us identify important gaps in our original submission and significantly improve the manuscript.

In response, we have substantially expanded our experimental evaluation and strengthened the supporting analysis:

  • We added empirical validation of the trajectory-level feasibility metric, together with analytical IK diagnostics and comparative placement evaluations.
  • We expanded the simulation benchmarks to 31 tasks across two simulators, three robot embodiments, and two VLA policies, with additional comparisons against recent baselines.
  • We extended real-world evaluation to closed-loop manipulation execution across six household tasks, providing further evidence of the method's physical applicability.
  • We added hyperparameter sensitivity studies and further clarified the method design, experimental settings, and key implementation details.

For a fully rendered version comparing our revisions against the initial submission, please refer to Revision_Summary.pdf in the Supplementary Material. In this document, newly added or substantially revised text is highlighted in blue, removed text is indicated in red strikethrough, and updated tables and figures are marked with blue revision tags in their captions. Videos of the real-world manipulation experiments are also included in the Supplementary Material (realworld_lastmeter_moma_video.zip).

We believe these additions provide stronger evidence for the effectiveness, generalizability, and practical applicability of StandReady. We sincerely appreciate the reviewers' valuable suggestions and hope that our revisions address their concerns.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.