acceptodds
Under review as a conference paper at ICLR 2027

Contextual Episodic Memory for Context-Aware Robot Navigation

Abstract

Embodied navigation in partially observable environments requires agents to retain information acquired earlier in an episode, since relevant objects, rooms, and landmarks may become inaccessible from the current observation. We propose Contextual Episodic Memory (CEM), a framework that progressively constructs episodic memory by integrating previously accumulated context into newly encountered observations. CEM selectively retains informative visual observations and associates each with the agent's 3D pose. Each new visual-spatial representation attends to stored episodic representations to retrieve relevant context, which is integrated with the current representation before being stored as the next memory entry. This produces an evolving representation of navigation experience that complements the current observation during decision making. We evaluate CEM in the photorealistic environment on two complementary tasks that require episodic memory in distinct ways: multi-room object search, where an agent must use information about previously observed objects and their locations after they are no longer visible, and target-room navigation, where an agent must retain scene-level experience gathered along the trajectory to navigate toward a room from varying starting locations. We compare CEM with recurrent, external-memory, and Transformer-based navigation approaches and analyze the contributions of informative observation selection and contextual memory construction. Results demonstrate that progressively constructed episodic representations improve navigation under partial observability by enabling agents to make decisions using experience accumulated throughout the episode.

Primary area:
applications to robotics, autonomy, planning
Keywords:
embodied navigation, episodic memory, contextual memory, visual memory, object search, long-horizon navigation, embodied AI, memory-augmented agents, spatial reasoning, AI2-THOR

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 4 reviews, above 0% of the others.
Reviewer_jtuxOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Contextual Episodic Memory (CEM) for visual navigation under partial observability. CEM retrieves context from previously stored visual-spatial experiences, fuses it with the current representation, and stores the resulting contextualized representation as a new memory entry. Experiments in AI2-THOR show improvements over recurrent and external-memory baselines.

Critical strengths
  1. The paper addresses an important problem in embodied navigation.

  2. CEM provides a simple and modular read–contextualize–write framework that combines online keyframe selection with explicit visual-spatial memory.

Critical weaknesses
  1. Limited novelty and insufficient justification of the proposed memory mechanism.

    • CEM primarily combines attention-based retrieval, simple linear feature fusion, and memory writing, without introducing a clearly distinctive memory mechanism.
    • Prior navigation methods such as SMT, VGM, and MemoNav have already explored incremental memory construction and contextual aggregation. The paper does not sufficiently explain its unique advantages over these approaches.
  2. Insufficient ablation of the central contribution.

    • The mixing coefficient affects both the policy representation and memory updates, making it difficult to attribute the gains specifically to contextualized memory writing.
  3. Limited experimental evaluation and insufficient comparisons with relevant navigation methods.

    • The baselines mainly consist of generic or relatively old architectures, such as LSTM, TrXL, DNC, and SAM, without comparisons to more recent navigation-specific memory approaches, such as 3D-Mem and MemoNav.
    • Evaluation is largely restricted to a limited AI2-THOR setting and its variants, without validation on established navigation benchmarks such as Gibson, HM3D, or Matterport3D.
  4. Poor writing quality and figure presentation.

    • The policy input is defined as z_t = [e_t; h_t], where e_t encodes only the current RGB observation and pose, without any target-image embedding. However, the paper reports target-image-conditioned navigation and experiments with 80% target-image occlusion, making it unclear how the target information affects policy decisions.
    • Table 6 reports 77.3% for Level-2 keyframes, whereas the accompanying text states that Level-2 alone achieves 70.0% and only Level-2+CEM reaches 77.3%.
    • The figures and overall layout have quality issues. Figure 1 suffers from poor clarity and readability, while Figure 2 appears to be AI-generated and lacks the quality expected of a carefully prepared ICLR submission.
Questions for authors

Please see weaknesses.

Reviewer_3hKxOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Contextual Episodic Memory (CEM), an episodic memory framework for robot visual navigation in partially observable environments. The method first selects representative keyframes from continuous visual observations and combines them with robot poses to construct visual–spatial features. It then retrieves historical information through an attention mechanism, integrates the retrieved context with current features using weighted fusion, and writes the contextualized representations back into memory to progressively accumulate context-aware episodic experiences. The authors conduct experiments in the AI2-THOR environment, including multi-room object search, target-room navigation, and object removal, and compare CEM with LSTM, DNC, SAM, and Transformer-XL. The results show that CEM achieves competitive navigation success rates and trajectory efficiency across various experimental settings.

Critical strengths
  1. The research problem is practically meaningful. This paper investigates how robots can leverage historical visual information to support navigation in partially observable environments. Episodic memory is well motivated for scenarios where targets are temporarily invisible or object locations must be inferred from previous observations.

  2. The framework is clearly structured and easy to integrate. The proposed method combines keyframe selection, visual–spatial encoding, memory retrieval, and policy learning in a straightforward architecture that can be readily incorporated into existing reinforcement learning-based navigation frameworks.

  3. The experiments cover diverse navigation conditions. The authors evaluate object search, object removal, navigation in unseen environments, and visual variations. CEM achieves favorable results on most metrics, while experiments on keyframe granularity and fusion coefficients provide some evidence of component effectiveness.

Critical weaknesses
  1. The methodological novelty is relatively limited, as the core mechanism mainly combines existing techniques. The proposed framework relies on established components, including keyframe selection, visual–pose feature fusion, scaled dot-product attention, and PPO. Its main methodological distinction is the fixed-weight linear fusion of retrieved historical information with current features, followed by writing the fused representation back into memory. However, updating current representations using historical context is already common in recurrent networks and attention-based memory architectures. The paper does not sufficiently clarify how this write-back mechanism fundamentally differs from existing memory update methods or what unique advantages it provides. Consequently, the independent methodological contribution appears limited.

  2. The core read–contextualize–write mechanism lacks a direct ablation, and inconsistencies exist in key experimental results. The central design of CEM is to write contextualized features back into memory. However, the existing ablations do not directly compare a configuration that uses historical attention for decision-making while storing raw current features against one that writes fused features back into memory. The fusion coefficient experiments mainly demonstrate the value of combining current information with historical context, but do not independently establish the necessity of the write-back operation. Furthermore, Table 6 reports an average success rate of 77.3% for Level-2, whereas the main text and appendix describe the success rate of Level-2 alone as 70.0% and attribute 77.3% to the complete CEM. This inconsistency between the table and textual descriptions makes it difficult to determine the actual contribution of the core mechanism.

  3. The target-conditioning mechanism of the navigation policy is insufficiently described. The paper specifies target objects in object search tasks and provides target images in target-room navigation. However, Section 3.4 and Appendix A.6 define the policy input as the concatenation of current visual–spatial features and contextualized memory features, without explicitly explaining how the target category or target image is encoded and provided to the Actor–Critic network. Although the reward function in the appendix incorporates target features, this does not substitute for a clear description of target conditioning in the policy input. As a result, the current methodological description does not adequately explain how the policy generates appropriate actions for different navigation targets.

  4. Performance improvements over strong memory baselines are modest, and some experimental conclusions require stronger support. Although CEM substantially outperforms memoryless policies and several recurrent baselines, its improvements over the more relevant SAM baseline are relatively small in several experiments. For example, the object search success rate improves by only 1 percentage point in the reference environment, the average success rate across four unseen environments improves by 2.4 percentage points, and the improvement under 80% target-image occlusion is only 0.64 percentage points. Some results lack variance estimates or confidence intervals across random seeds, making it difficult to assess the consistency of these gains. Furthermore, the paper does not sufficiently report memory size, retrieval latency, or computational overhead, limiting a comprehensive assessment of the practical efficiency of the proposed memory construction mechanism.

Questions for authors
  1. Could the authors include a direct ablation comparing the storage of raw features against the storage of contextualized features, while keeping keyframe selection, attention-based retrieval, policy inputs, and training conditions identical? Such an experiment is essential for isolating the contribution of the proposed write-back mechanism.

  2. Could the authors clarify the discrepancy between the 77.3% success rate for Level-2 in Table 6 and the 70.0% reported in the main text and appendix? What specific configurations do Level-2 and Level-2+CEM represent? Does the current table omit a baseline without contextualized write-back?

  3. How are target categories or target images provided to the policy network in navigation tasks involving different target objects or rooms? Please clarify whether the target representation is included in ete_t, hth_t, or another unspecified input, and whether separate policies are trained for different targets.

Reviewer_xLaeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes CEM, which stores the images and poses observed by the agent as latent representations. It then retrieves historical representations through an attention mechanism and concatenates them with the current context to perform navigation. The latent representations are subsequently written into persistent memory. The paper conducts multi-object search and generalization experiments in the AI2-THOR environment and achieves relatively good results compared with the authors' baselines. However, the experimental setup is not transparent, and the comparison methods are insufficient, making the actual effectiveness of the experimental results questionable.

Critical strengths
  1. The paper proposes CEM, which stores memory as vector representations at the set level. This is a good idea, but the paper has significant shortcomings.
Critical weaknesses
  1. The main framework completely lacks the incorporation of target information. The Method section does not mention it at all. How exactly does the model receive target information?

  2. In the experimental section, only the simulation environment is publicly specified. The paper does not mention how many trajectories were used or how the trajectory dataset was split, which makes me highly doubtful of the experimental results. Meanwhile, the baselines in all tables lack corresponding citations, and the paper does not provide details on how they were applied to its own framework. I do not consider these acceptable experimental comparisons. Furthermore, the paper does not compare its method with recent memory-based VLN methods. The lack of transparency regarding the experimental dataset and the absence of comparisons with contemporaneous papers make the paper's claims questionable.

Questions for authors

None.

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

CEM selects visual keyframes, combines each with planar pose, and retrieves context from an external episodic memory using attention. A fixed mixture of the current representation and retrieved context is used by a PPO policy and stored as a detached memory entry. Experiments cover object search, revisiting locations after object removal, and target-room navigation in AI2-THOR, with comparisons to recurrent, associative-memory, and Transformer-based policies.

Critical strengths
  1. Writing contextualized experience rather than storing every raw observation is a simple and potentially useful memory design. Object-removal evaluation is particularly relevant because the current target cue is absent. The reported recall is 65.89%, versus 60.24% for SAM and 56.75% for DNC (Table 3), providing a more targeted memory test than standard navigation success alone.
  2. The evaluation includes several perturbation settings and retains competitive baselines rather than only a no-memory comparison. CEM reaches 64.0% versus SAM's 58.0% on the duplicate-object setting, with five-seed variation reported in Table 1. On four unseen room-navigation environments, mean success is 78.94% versus 76.54%, accompanied by shorter reported paths. These results motivate further investigation of the complete navigation system.
Critical weaknesses
  1. The main contribution requires a matched contextual-write versus raw-write control. The alpha=0 variant removes memory context from the policy as well as from newly written entries: its input becomes two copies of the current embedding. It therefore cannot isolate the value of storing contextualized representations. Table 6 states that contextual mixing is fixed and reports Level-2 at 77.3%, whereas Section 5.4 and the conclusion claim a separate 70.0% Level-2 baseline and a 77.3% Level-2+CEM result; the 70.0% configuration has no corresponding table row. The memory-granularity experiment is consequently not a reproducible test of the novel write-back step. More broadly, the contrast with recurrent/contextual memory is overstated unless the benefit survives a control with the same keyframes, pose inputs, read operation, and policy capacity.
  2. The task and implementation descriptions leave consequential ambiguities. The target-room policy is said to receive a target image, but Sections 3.2-3.4 and Appendix A.6 define its entire input as current RGB/pose plus memory derived from those inputs; the goal embedding appears in the reward without a specified policy-conditioning path. Object search alternates between a single specified target and discovering all five objects before termination. The keyframe procedure on pp.12-13 uses centroid selection and a three-term utility, while Appendix A.4 uses BLIP/spaCy grouping and a different five-term quality score. Finally, object-removal evaluation does not specify the within-episode observation, removal, query, and memory-reset schedule. Without these details, it is unclear which actual task and memory construction produced each table, and whether removal recall distinguishes episodic information from a learned location prior.
Questions for authors
  1. Please report a comparison that changes only the stored vector from the contextualized representation to the raw visual-spatial embedding, while preserving memory retrieval and the policy input. Which exact configuration produces the claimed 70.0% result, and how does it correspond to Table 6?
  2. Where and how does the target image or object identity enter the actor and critic? Please specify the single-target versus five-target protocols, success predicates, episode counts, and the observation/removal/query/reset timeline for Table 3. Were target locations varied independently across episodes?
  3. Which of the two keyframe-selection algorithms generated the experiments, and did all memory baselines use the same selected observations, pose access, reward, training budget, and memory capacity? Please also clarify whether path-length averages include failed episodes, since shorter trajectories alone do not establish success-normalized navigation efficiency.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.