acceptodds
Under review as a conference paper at ICLR 2027

Load-Aware Quadruped Navigation with Contact-based Online Geometry Estimation

Abstract

Navigating cluttered environments while carrying objects remains a challenging problem for quadruped robots. Compared with extensive works that focused on load-aware locomotion adaptation to maintain gait stability, existing autonomous navigation systems generally ignore the influence of carried objects' geometries and use fixed footprint inflation, leading to overly conservative navigation. To address this, we propose a two-phase load-aware navigation learning framework. In Phase 1, we train LOADNav, a load-aware navigation policy via reinforcement learning, enabling the robot to adapt its trajectory and clearance behavior according to the geometric size of the carried object. In Phase 2, we train CoGE, an online geometry predictor via imitation learning, which estimates unknown payload geometry from occasional contact events. We evaluate the framework in cluttered environments with loads of varying sizes and shapes. The results show that our navigation policy exhibits clear load-aware behaviors, and the online geometry predictor estimates object geometry from occasional contact events, refining its estimate and adapting the navigation behavior accordingly. Without requiring any prior knowledge of the payload, it narrows the gap to geometry-aware policies.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Quadruped Robots, Load-Aware Navigation, Robot Navigation, Online Geometry Estimation, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_R1UAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a two-stage framework to address payload-induced sensor blind spots and changes in geometric constraints during quadruped navigation. In the first stage, a LOADNav policy is trained through reinforcement learning to adapt its trajectory and orientation to the payload geometry. In the second stage, a proprioception-based geometry predictor estimates the dimensions of an unknown payload. A recurrent memory mechanism is further introduced to mitigate the loss of contact information caused by a finite observation window, enabling payload-size estimation without prior geometric knowledge.

Critical strengths
  1. The problem is practical and distinct from ordinary obstacle avoidance. Standard navigation treats the robot as a rigid body of fixed size, whereas the obstacle-avoiding body is the robot together with its payload.
  2. The experiments are reasonably thorough, with easy-to-interpret scenarios covering payloads of different geometries and validation on a real robot.
  3. The observation that payload geometry is observable only through contact is insightful. The added model addresses a concrete observability limitation rather than increasing complexity arbitrarily.
Critical weaknesses
  1. The way ACT is computed introduces a clear selection bias. Because ACT is averaged only over successful trials, a method with a higher success rate may retain more trajectories that succeed after collisions and consequently report a higher ACT. Therefore, ACT should not be interpreted as a standalone safety metric.
  2. The analysis relies on several idealized assumptions, including that the payload is rigidly attached and that its dynamic effects during contact-free motion are negligible. These assumptions may limit the method’s applicability to more realistic transportation tasks.
Questions for authors
  1. Could the authors report the number of collisions over all attempted trajectories, the success rate among trials with no payload contact, and the fraction of trials that become stuck and fail after a collision? Does CoGE-LOADNav+ actually make more contacts and simply recover from them more easily than CoGE-LOADNav?
  2. Before the first informative contact event, how does the system handle uncertainty about the unknown payload dimensions? Does the policy adopt a more conservative route, or can underestimating the payload size cause it to select a passage that is not actually traversable?
Reviewer_LDPGOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper deals with the challenge of navigating in cluttered environments while carrying objects, where the object size would affect the navigation trajectory. The paper also assumes that the object size is unknown to the system and can be estimated from occasional contact (collision) events. The authors first train a navigation policy with PPO that conditions on the privileged size information, then use the privileged policy as a teacher to learn an online predictor that estimates the object size with recurrent memory.

Critical strengths

Part of the target problem, i.e., load-aware navigation in cluttered environments, is an interesting and useful novel task, and the paper provide an effective solution that also works in the real world.

The paper, especially the first three sections, are clearly written and the illustrations are nice and easy to understand.

Critical weaknesses

One of my biggest concerns is about how practical the assumed setting is. CoGE-LOADNav assumes the loaded object’s size is unknown to the robot, but the object can undertake collisions that enable the robot to estimate the size. First, because the payload geometry is defined by two parameters, sx and sy, I don’t think providing the size to the system would be very costly so that the ability to deal with unknown object would be valuable. Second, the applications where collisions are allowed, even though occasionally, would be very limited in my opinion. So, at least half of the paper, i.e., CoGE-LOADNav, is trying to solve a problem that is not well motivated.

Even if the target problem is considered useful and interesting, the proposed solution is consisted of mature methods, i.e., PPO, privileged learning and recurrent state estimation. Although applying these methods on a novel problem is not a trivial effort, I would expect more novel contribution in terms of scientific insights and methodology for a paper accepted by a top ML venue.

In addition, the evaluation metrics only include success rate and collision times, but an important aspect of the motivation of the target problem is that the robot should be able to choose the best route according to the payload geometry, which cannot be reflected by the two metrics. A navigator that always choose the most conservative route will get high SR and low ACT.

Lastly, it seems that the methods are evaluated on very limited number of scenes, i.e., only one layout for each of the densities in the simulation and one for the benchmark and one for the real-world experiments. Therefore it’s hard to see if the methods are sensitive to the layouts and if so, the current results could be invalid.

Minor issues:

  • The naming of “the simulation” and “the benchmark” is confusing. “The simulation” is also used as a benchmark in Table 7 and “the benchmark” is also simulated.

  • L148: sxs_x or sxs^x?

  • Video 01:49-01:52: it should be Tube_X and LOADNav?

Questions for authors

What do you think of the practicality of the assumed setting? In what scenarios is the CoGE-LOADNav particularly more useful than LOADNav and how common would they be?

How would the paper be useful for a broad audience that doesn't specifically care about the load-aware quadruped navigation problem?

How can you quantitatively evaluate whether the policy choose the route based on the object size?

Reviewer_wB1COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a two-phase load-aware navigation framework for quadruped robots carrying objects with unknown geometry. The paper aims to address the limitations of fixed-footprint navigation when payloads enlarge or reshape the robot's effective boundary, particularly when their geometry cannot be directly observed by onboard sensors. To tackle these issues, the authors first train LOADNav, a reinforcement learning policy conditioned on payload dimensions, proprioceptive observations, and a local occupancy map, enabling geometry-dependent trajectory and orientation selection. The navigation policy is subsequently frozen, and a contact-based online geometry estimator, CoGE, is trained using geometry regression and action imitation to infer payload dimensions from proprioceptive history and occasional contact feedback. In addition, the paper introduces a recurrent estimator that retains intermittent contact information beyond a finite observation window, mitigating the forgetting of previously inferred geometry. Experiments across different payload shapes and cluttered environments demonstrate load-aware navigation behaviors and show that online estimation reduces the performance gap to policies supplied with known payload dimensions.

Critical strengths
  1. The paper tackles a useful geometric adaptation problem with clear evidence that the geometry input matters. Payloads with different orientations require different clearance and rotation strategies even when the base robot and goal are unchanged. In the warehouse comparison, LOADNav reaches 91.67% success versus 63.33% for the size-blind LOADNav-B, and the Board Y case drops from 85% to 45% without geometry input. This control is particularly helpful because it connects the proposed conditioning to navigation performance rather than relying solely on comparisons to unrelated methods.

  2. The online estimator is integrated into a functioning navigation system and evaluated on hardware. The recurrent variant addresses an observed failure mode in which contact information leaves the finite observation window. Its reported geometry and action MSE reductions, trajectory examples, and real-world success of 90% compared with 81.67% for the finite-memory variant provide complementary evidence. Although privileged training, distillation, and recurrent memory are established components, their application to intermittent geometric feedback is a useful robotics contribution supported by physical experiments.

Critical weaknesses
  1. The claim that geometry is informative only at contact is stronger than the proof supports. Geometry-independent nominal dynamics do not exclude information from contact-free traversal or correlations between payload dimensions. Equation (6) and the finite-window forgetting claim therefore require clearer assumptions about the available history, limiting the strength of the theoretical contribution.

  2. The evaluation does not fully establish the cost or reliability of contact-based adaptation. ACT uses inconsistent denominators across experiments, and excluding failed trials obscures overall collision exposure. Small deployment comparisons lack uncertainty estimates, while the inflation baseline does not account for orientation-dependent geometry. These gaps limit conclusions about robustness and comparative performance.

  3. Methodological novelty is limited. Privileged-geometry training followed by estimator learning with a frozen policy closely follows established teacher-student adaptation, and recurrent memory is a standard extension. The application to intermittent payload-contact feedback is useful, but the paper does not clearly establish a substantial methodological advance beyond adapting these components to the task.

Questions for authors
  1. Could the authors clarify the scope of the identifiability argument in Section 3.4? Showing that nominal motion is independent of payload geometry without contact does not necessarily imply that an entire contact-free history is uninformative. Traversing a known clearance without contact may constrain payload size, while correlations between the two dimensions may affect the prior-variance bound in Equation (6). Please specify the estimator's available information and the assumptions required for this bound. The finite-window forgetting claim would also benefit from distinguishing the absence of direct contact feedback from the absence of geometric information in the observation-action history.

  2. The collision statistics are difficult to compare across experiments because their denominators differ. Section 4.1 defines ACT over successful trials, whereas Table 7 reports an average over all trials. The explanation in Section 4.4 further indicates that excluding failed runs can alter the comparison between the finite-memory and recurrent variants. A consistent report of collision counts across all attempted trials, alongside the fraction of successful runs without payload contact, would help determine the trade-off between navigation success and contact exposure. How is a collision counted: as a new contact event or as repeated samples during sustained contact? Please also provide uncertainty estimates for success rates and, if recorded, contact duration or force statistics.

  3. More detail is needed to understand what enables CoGE-LOADNav+ to retain geometry estimates and how reliably it does so. In particular, please describe the input channels, initial estimate, recurrent architecture, hidden-state reset rules, sequence length, and training rollout collection procedure. Command-tracking discrepancies can also result from slipping, external pushes, or low-level control errors. How does the estimator respond to these disturbances when no payload contact occurs? Existing results or diagnostic examples would help establish whether the learned update reflects payload geometry or more general motion deviations.

  4. What methodological advance does the proposed framework offer beyond adapting an existing teacher-student approach to this task? Training with privileged geometry, freezing the policy, and learning an estimator through geometry regression and action imitation closely follow the two-phase adaptation paradigm discussed in Section 2. Recurrent memory is likewise an established way to retain intermittent observations. Please identify which design choices are specifically motivated by contact-based geometry estimation and provide evidence of their benefit over a standard adaptation model using the same inputs and low-level controller. The baseline protocol also needs clarification: were VP-Nav and NavRL adapted and trained under comparable conditions, and was an orientation-aware footprint planner with known payload dimensions evaluated? These comparisons would help isolate the contribution of the learning design from the value of geometry information itself.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.