acceptodds
Under review as a conference paper at ICLR 2027

OC-RAP: Observation-Consistent Recovery-Affordance Planner

Abstract

Safety-critical planning does not end at avoiding first contact. Near the safety boundary, a planner should preserve recoverable continuations; after contact, it should separate, avoid secondary events, and return to a viable state. Under partial observation, however, branch-wise planning can be optimistic: each latent future may admit a recovery while observationally indistinguishable futures require incompatible recovery actions. We formalize this mismatch as an oracle-to-deployable recoverability gap and introduce OC-RAP. For each candidate prefix, OC-RAP rolls out actuator-realizable recovery options under latent roots. Its core operator, OC-MERO, lower-tail aggregates each fixed option over roots compatible with the same post-prefix observation before option selection, preventing hidden root identity from authorizing an unrealizable contingent recovery. A signed margin represents positive recovery reserve and negative recovery debt, providing one recovery semantics before and after contact, while RIFA admits non-nominal actions only when conservative absolute recovery evidence is sufficient and otherwise preserves the exact nominal action. On frozen real-world-motion cohorts executed in closed loop, OC-RAP has zero Safe intervention with nominal-utility preservation (NUP) . On 250 Near-Contact scenes it obtains the best non-collision clearance p05 (0.661 m), TTC p05 (0.130 s), and TTC-deficit AUC (0.630 s), although the best comparator has a lower collision rate and another has shorter critical-TTC exposure. On Contact, OC-RAP attains the shortest overlap, lowest penetration AUC, and largest terminal clearance on the fully paired 8-anchor physical comparison; a complementary continuous re-contact audit gives the highest initial-separation rate (88.5%) and the lowest post-separation re-contact rate (25.0%) and exposure (8.6%). Matched ablations support the roles of observation consistency, executable recovery geometry, absolute admission, and exact-nominal abstention.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Autonomous Driving, Motion Planning, Partial Observability, Safety-Critical Planning, Recovery Planning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ZMhWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies recovery planning under partial observation, where a planner may appear recoverable when evaluated with hidden future information, but the selected recovery strategy may not be executable during real deployment. The authors identify this issue as the oracle-to-deployable recoverability gap and propose oc-rap, an observation-consistent recovery-affordance planner.

The core component, oc-mero, changes the recovery selection procedure by aggregating observation-compatible latent futures before selecting a recovery option, preventing the planner from relying on unavailable future information.

The paper also introduces a signed recovery margin to unify recovery reasoning before and after contact events, together with rifa to separate recovery evidence from intervention preference. Experiments on Waymo Open Motion Dataset with Waymax closed-loop evaluation demonstrate improvements in several Near-Contact and Contact scenarios while maintaining conservative behavior in Safe scenarios.

Critical strengths

The main strength of this paper is the identification of the oracle-to-deployable recoverability gap. The paper highlights an important issue that recovery planning methods may implicitly use information that is unavailable during deployment. This provides a clear motivation for enforcing observation consistency in safety-critical decision making.

The proposed oc-mero operator is well aligned with the identified problem. Instead of selecting recovery actions separately for each latent future and then aggregating the results, the method first evaluates a fixed recovery option across compatible futures and only then performs selection. This design directly addresses the information constraint considered in the paper.

The experimental evaluation is carefully designed. The authors evaluate Safe, Near-Contact, and Contact settings using a unified framework and provide component-wise ablations. The results suggest that the proposed observation-consistent aggregation and rifa mechanism contribute to improved recovery behavior rather than only relying on a single metric improvement.

If the authors can better establish the novelty of oc-mero compared with existing non-anticipative planning methods and provide stronger evidence about the practical validity of the soft compatibility approximation, I would consider increasing the rating.

Critical weaknesses

The main concern is the novelty of the proposed method. The paper acknowledges that the general principle of making decisions based only on available information is not new. The connection between oc-mero and existing non-anticipative planning, stochastic optimization, and risk-sensitive decision methods needs to be clarified more carefully. Without a stronger distinction, the contribution may appear as a domain-specific adaptation of existing ideas.

There is a gap between the theoretical analysis and practical implementation. The theoretical guarantee is established under hard observation partitions, while the actual method uses a soft compatibility kernel as an approximation. It is unclear how robust the proposed method is when the compatibility estimation contains errors or when different latent futures are incorrectly grouped.

The method relies on several learned or predefined components, including latent future generation and recovery option libraries. If the latent future model misses important failure cases or the recovery set does not contain suitable actions, the benefits of oc-rap may be limited. Furthermore, the physical Contact evaluation is based on a relatively small number of paired scenarios, which limits the strength of the generalization claim.

The evaluated final selector gives nonzero weight 0.05 to distance from the logged future ego trajectory. That future is unavailable to a deployed observation-only planner. Without the authors' proposed zero-weight rerun, the reported closed-loop numbers cannot establish the abstract's full deployability claim. This does not refute the OC-MERO operator or its hard-partition result.

Questions for authors
  1. How does oc-mero fundamentally differ from existing non-anticipative stochastic planning or risk-aware contingency planning methods? Could the authors clarify which part of the formulation cannot be directly derived from existing frameworks?

2.The theoretical analysis assumes hard observation partitions, while the practical implementation uses a soft compatibility kernel. How sensitive is the performance to errors in the compatibility estimation?

3.How sensitive is oc-rap to errors in the latent future generation process? If the latent roots fail to cover important failure modes, how does this affect recovery reliability?

  1. The recovery option library appears to play an important role in determining available recovery behaviors. How does the method perform when the optimal recovery trajectory is not included in the candidate recovery set?
Reviewer_VLsgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose an autonomous driving planner that focuses on maintaining performance even after a collision. Despite significant effort, I was unable to understand the proposed approach at all, much less the claimed contribution. Dense but non-standard jargon and unexplained notation permeate the paper.

As far as I can tell, the proposed approach is an ad-hoc planning method that simulates several candidate trajectories, choosing the best ones based on how well the trajectory might perform even after a collision event. It is unclear whether or not there is any learned component, or whether the approach is at all reasonable.

Critical strengths

None

Critical weaknesses

The proposed approach may be valuable, but the manuscript is simply unreadable in its current form. See the questions below for a detailed breakdown of these readability issues.

Questions for authors

The AI use statement acknowledges that "Generative AI tools were used to improve manuscript readability" -- which AI tools, and what was the prompt? Because this really did not work well at all. The manuscript needs a complete rewrite.

In lieu of questions, which I don't have many of because I wasn't able to make sense of the manuscript, I'll use this space to provide specific comments on readability issues. Hopefully this helps the authors produce clearer writing in the future.

  • The abstract needs to define the problem at hand -- "initial contact" with what? "Safety-critical planning" for what? Only at the end of the first page do we learn that this paper is about autonomous driving.
  • Acronyms should generally be avoided in the abstract.
  • Benchmarks that are specific to this paper/project, like "Safe" and "Contact", should be defined before they are referred to. Referring to such non-standard things is a hallmark of AI slop.
  • Use language that is as simple as possible, but no simpler. Jargony language and complex sentence structure should only exist when absolutely necessary to support a complex idea. Sentences like "each latent future may admit a recovery even when observationally indistinguishable futures require incompatible recovery actions" are the opposite of this. I spent about 10 minutes trying to parse this sentence, and in the end I'm not sure it even makes sense. What is a "future"? What makes one "indistinguishable"? In an MPC setting, which I'm guessing we're in with "futures" = predictions, you replan at each step, so it doesn't matter if the "branch"/prediction the controller has selected at one step turns out to be suboptimal: the "branch" can and will change based on observations during operation.
  • "otherwise the planner preserves the exact nominal action" - I'm not sure what this means exactly, but this sounds very dangerous for autonomous driving.
  • What makes the futures "latent"? Is a world model involved, where the system makes predictions in some learned latent state? If so this should be clearly stated.
  • "After contact [the vehicle should] return to a viable route state" - I don't agree with this premise. I think I would prefer if my autonomous vehicle just stopped after a collision!
  • Many acronymed components are introduced: OC-RAP, OC-MERA, RIFA, etc. Are these standard methods, different approaches, or components of the same approach? Their relationship is unclear.
  • A simple overview diagram might help. Figure 1 was unreadably small when I printed the paper to make notes and prepare this review.
  • What does it mean that "no single option is valid across branches that are still observationally aliased"? Again, predictive control methods do not require every prediction to have the same actions - quite the opposite in fact! "Observationally aliased" sounds like some AI-hallucinated jargon; it's not a term of art that I have ever heard.
  • The problem formulation introduces a fair amount of notation, but it is not clearly defined. What sort of mathematical object is the available history, prefix, latent root, etc.? What does the "latent root zkz_k" represent? Is it something you learn? The system state?
  • What is the "hidden root oracle"? What does this oracle provide? What is V? A value function? Something else?
  • Mixed in with the excessively complex notation are some very obvious facts, like referring to probabilities and then specifying that they must sum to one. Such a mix of extremely complex, ill-defined notation and overexplained obvious facts is another hallmark of AI slop.
  • The terminology of "surplus" and "margin" is highly nonstandard. It took me a long time to realise that this just meant "constraint violation."
  • It seems that the proposed approach is premised on a discrete action set, but autonomous driving deals with fundamentally continuous inputs.
  • "the implementation also reports a privileged choose-first oracle diagnostic" - What does this mean? Is this part of the proposed approach, or a little extra bonus in your particular implementation? Again I fail to follow the "oracle" terminology.
  • "privileged truth-side values" - what does this mean?
  • In Section 5 (experiments), we suddenly hear about a "learned recovery model." How is this model trained? Is it part of the contribution in this work?
  • Much of the text in section 5 is devoted to restating results that are already listed in tables. Use the text to highlight trends and conclusions, rather than presenting exactly the same data twice.
  • Throughout the entire text, nearly every paragraph contains phrases that are, to the best of my knowledge, entirely new and made up jargon. Examples include "one-factor inference-time knockouts," "non-compensatory signed margin," "pre-frozen scalar recovery-admission threshold," "logged-future route flag," etc. These terms are never defined or explained, which is the major issue preventing readability in the manuscript.
Reviewer_o6GuOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

A major weakness of this paper is its clarity, which strongly influenced my weak understanding of the paper. It seems the paper deals with controlling some sort of autonomous vehicle.

There is a distribution over the current state (the "roots") and a set of action sequences to choose from (the "prefixes"). Given a root and a prefix, we can compute a set of possible futures (the "latent futures"), and for every such future there is a possible recovery policy chosen from a finite recovery library.

Now, this combination is associated with a cost of some sort, and the objective is to choose prefixes that minimizes this cost. An additional "twist" is that after performing the prefix, and before choosing a recovery policy, the vehicle gets a "post-prefix observation", which are related to the actual state after performing the prefix, and helps the agent decide better which recovery policy to use.

The main "novelty" is that choosing the best prefix is not the same as choosing the best prefix for every possible future, since in the current state we do not know which future will be chosen.

The proposed system, called OC-RAP, seem to help selecting appropriate policies and recovery policies. They evaluate on data from Waymo adapted to their setting, and compared against several baselines.

Critical strengths
  1. The problem seems to be very practical and challenging.
  2. The experimental evaluation based on real data is commendable.
Critical weaknesses

As I wrote above, major weakness is readability. The paper contains many terms that are ambiguous and unclear to me. Here are some examples:

  • "branch-wise planning"?
  • "latent future"?
  • "post-prefix observation"?
  • "hard-partition limit"
  • "observation-measurable"
  • "signed recovery quantity"
  • "latent-root identity"
  • "signed root–option margins"

In general, I found the terminology somewhat non-standard. If I understood correctly, what they call a "root" is the current state, a "prefix" is a sequence of action or a policy they want the agent to perform, and a "latent future" is the possible sequence of states the agent may reach after performing a sequence of actions from the current state. I recommend that the authors use terminology from the POMDP literature, or some other well-founded literature. This will not only make the paper more readable, but also allow better mapping to prior works.

I was also disappointed that the problem definition section does not fully define the problem, i.e., what problem they wish to solve and what do they want to optimize. There is a hint of this in line 161 of page 3.

What I think would help is a major edit of the text, along with very concrete pseudo-code and problem definition section.

Minors:

  • In "frozen real-world-motion" and "Frozen closed-loop evaluation", what makes them "frozen"? what is a not frozen real world motion or closed loop evaluation?
  • "On Contact, OC-RAP attains..." - very long sentence, hard to parse.
  • "... executable recovery geometry, absolute admission, and exact-nominal abstention" - I do not know what these term mean in this context without further explanation.
  • Line 182: "root-option" is there with a wrong type of dash (in latex, it should be "-" and not "--")
  • What is the purpose of the lower-tail functional in line 145 in page 3? What does it represent? Is it assumed to be given as input to the algorithm?
Questions for authors

Q1. You write as the first contribution that you "oracle-to-deployable gap induced by partial observation". However, as you wrote, this gap is well-known: "the novelty is not the general principle that decisions may depend only on available information (Shapiro et al., 2021; Kaelbling et al., 1998), but ..."

Q2. I am not sure the listed related work on POMDP do not include a "safe option may exist in every hidden branch while no single option is valid across branches that are still observationally aliased". In particular, conformant planning does exactly that: searching for a plan that will work regardless of the outcome given the current belief. So, how is your work different? can we take conformant planners to solve your problem? if not, why?

Comments on the paper

1 comment
Reviewer_ZMhWOpenReview ↗

AI Use Clarification

The text bellow is my original, human-written review draft, reproduced exactly as I submitted it to the AI tool, including any grammatical or spelling errors. I used AI to polish the language and check for potential factual errors, inconsistencies, and unclear expressions. The substantive assessment, strengths, weaknesses, questions, and final rating were independently determined by me. The original text is provided here to clarify the scope of AI assistance and demonstrate my independent evaluation of the paper.

Summary This paper study recovery planning under partial observation. Planners can look recoverable with hidden future info, but strategies fail in real use. Authors call this the oracle-to-deployable recoverability gap, and propose OC-RAP. Core part OC-MERO aggregates observation-compatible latent futures before picking recovery options, so planner don’t rely on unavailable future info. Paper also uses signed recovery margin and RIFA to seperate recovery evidence from intervention preference. Waymo dataset tests show better near-contact/contact performance, stays conservative in safe cases.

Strengths

  • Good catch on the oracle-to-deployable gap. Highlights recovery methods may use undeployable info, clear motivation for observation consistency.
  • OC-MERO fits the problem well: aggregate compatible futures first then select, directly targets the info constraint.
  • Well-designed experiments with ablations across 3 settings. Gains come from both OC aggregation and RIFA, not just single metric.
  • May raise rating if novelty over non-anticipative methods is better established, and more proof for soft compatibility approximation.

Critical Weaknesses

  • Novelty not fully convincing. “Only use available info” is not a new principle. Need clearer distinction from existing non-anticipative / stochastic planning work.
  • Theory-implementation gap. Theory uses hard observation partitions, real method uses soft compatibility kernel. Unclear robustness when compatability estimation has errors.
  • Relies on latent future generation and recovery libraries. Missing failure cases or bad recovery sets limit benefits. Few contact scenerios also weaken generalization claim.
  • Final selector weights logged future trajectory (0.05), which is unavailable in deployment. Results don’t fully support deployability claim without zero-weight rerun.

Questions

  1. How is OC-MERO fundamentally different from existing non-anticipative stochastic / risk-aware planning methods? What part can’t be derived from existing frameworks?
  2. How sensitive is performance to compatibility estimation errors, given theory uses hard partitions but implementation is soft?
  3. How does latent future generation error affect OC-RAP? If key failure modes are missed, what happens to recovery reliability?
  4. How does the method work when optimal recovery trajectory is not in the candidate library?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.