acceptodds
Under review as a conference paper at ICLR 2027

MSPR: Multi-scale Predictive Representations for Goal-conditioned Reinforcement Learning

Abstract

This paper investigates robust representation learning in offline goal-conditioned reinforcement learning (GCRL). Particularly in sparse reward scenarios, learning representations that align state and goal latents is a challenge, as the encoder can learn goal-agnostic features that destabilize policy learning. We address this issue by learning the encoder's representation with alignment objectives that capture the environment across multiple scales, from local physical dynamics to long-horizon goal-directed structure. Concretely, we propose MSPR, a framework that leverages multi-scale predictive supervision to enforce goal-directed alignment within the latent space. We demonstrate that MSPR leads to strong performance on both vision and state-based tasks. Furthermore, we show that our approach is resilient under realistic, challenging data regimes, maintaining state-of-the-art performance across a wide variety of tasks.

Primary area:
reinforcement learning
Keywords:
Representation Learning, Offline Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_ccxiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper introduces multi-scale predictive representations, which try to predict different properties of offline behavior trajectories with the goal of finding a representation suitable for downstream offline goal-conditioned RL. The paper proposes to combine different alignment objectives in order to learn a good representation. Namely, the proposed goal representations is based on dynamical (next-state prediction), temporal (return prediction) and behavioral (inverse model, goal dynamics prediction) alignment. The authors give intuition why these alignment objectives might be helpful for offline GCRL and show the advantage of their representation empirically on OGBench. Ablation studies underlie the importance of each representation objective, showing that the required alignment objectives are task specific.

Critical strengths
  1. Combining different representation objectives in order to structure the representation allows for a general approach that is agnostic to downstream RL methods
  2. Convincing empirical results. The approach beats multiple baselines on a variety of OGBench tasks. The additional analysis of failure cases shows limitations of the representation learning approach itself. The ablations show that each objective has task-specific benefits, indicating that a combination of alignment objectives is required to solve different tasks.
  3. The paper has a good structure and is well written
Critical weaknesses
  • W1: While intuition is given for the different alignment objectives, an overarching principle is missing and it remains unclear which of these objectives contribute to the final performance or are redundant. For instance, the dynamic losses were never ablated in total isolation (i.e. no dynamics at all). In summary, the approach feels like a combination of different representation objectives without a clear overall picture.
  • W2: The dynamics losses and latent unrolling is closely related to model-based RL while the approach is only compared against model-free GCRL approaches.
  • W3: Novelty unclear. The architecture and objectives are quite similar to MR.Q (as acknowledged by the authors) and extend the approach to the GCRL setting with additional behavioral alignment.
  • W4: The idea should be agnostic to the downstream GCRL approach as stated by the authors, however only TD3+BC as backbone is considered.
Questions for authors
  • Q1: How important are the dynamics losses? Especially as they resemble a latent world model (W3) and are not ablated fully (W2). An additional ablation without any dynamic losses could be helpful.
  • Q2: What is the contribution of this paper? The motivation of the different objectives is not clear (W1) and the approach seems very close to MR.Q (W3).
  • Q3: Is the approach agnostic to the down-stream GCRL approach? Could the representation be used with e.g. CRL or GCIQL?
5 comments
AuthorsOpenReview ↗

(1/3) Response to Reviewer ccxi

We thank the reviewer for the careful review. We answer each point below. Changes in the revised paper are shown in blue.

W1. "An overarching principle is missing and it remains unclear which of these objectives contribute to the final performance or are redundant."

The principle is simple. The MSPR encoder gets no gradient from the actor or the critic. So whatever the agent needs must already be in the representation, and each loss puts one piece of it there:

  • Lrew\mathcal L_{rew} (Monte-Carlo return): how far the goal is. The critic reads it from (zsa,zg)(z_{sa}, z_g).
  • Ldyn\mathcal L_{dyn}: the next state. The critic bootstraps from it.
  • Lg-act\mathcal L_{g\text{-}act} and Lg-dyn\mathcal L_{g\text{-}dyn}: the goal-directed action and where it leads. The actor reads them from (zs,zg)(z_s, z_g).

Propositions 1 to 3 bound the error that each missing piece can cause, and none of the three error terms controls another (Sec. 4, "The three terms are distinct"). So in the analysis no condition is redundant. The ablation agrees (selected tasks; all tasks in Table 2):

Environmentw/o Ldyn\mathcal L_{dyn}w/o Lrew\mathcal L_{rew}w/o Lg-dyn\mathcal L_{g\text{-}dyn}w/o Lg-act\mathcal L_{g\text{-}act}MSPR
antmaze-giant540575568
humanoidmaze-medium56007567
humanoidmaze-large53004358
pointmaze-large2431623348
cube-single5063686265
visual-puzzle-3x3213902530
Avg. state (12)5947586367
Avg. pixel (5)6458556668
  • Without Lrew\mathcal L_{rew}, the encoder receives no direct supervision for goal-conditioned returns, leaving the critic to learn goal proximity through sparse-reward TD bootstrapping. On antmaze-giant and both humanoidmaze tasks, success drops to 0.
  • Without Lg-dyn\mathcal L_{g\text{-}dyn}, the actor loses the goal-directed signal, and both humanoidmaze tasks and visual-puzzle-3x3 drop to 0.
  • Without Ldyn\mathcal L_{dyn}, the drop is spread across tasks. It is largest on pointmaze-large and cube-single, where the next state depends on small details (a wall, the contact with the cube).
  • Removing any one loss lowers both the state and the pixel average.

Sec. 5 (RQ1) now explains each drop by the role of the loss.

Q1. "How important are the dynamics losses? … An additional ablation without any dynamic losses could be helpful."

We ran it. No-dyn removes both Ldyn\mathcal L_{dyn} and Lg-dyn\mathcal L_{g\text{-}dyn} and keeps the other two losses. For reference, No-aux removes all four losses (same network, encoder trained by the critic's TD loss). Same agent and hyperparameters (Table 8, App. E.4):

EnvironmentNo-auxNo-dynMSPR
pointmaze-medium503454
pointmaze-large422348
antmaze-medium09799
antmaze-large09098
antmaze-giant03268
humanoidmaze-medium24467
humanoidmaze-large0058
cube-single535265
cube-double151713
scene383245
puzzle-3x3899898
puzzle-4x4838393
Avg. (12)334767

The dynamics losses matter most on long horizons. Without them, antmaze-giant drops from 68 to 32, and both humanoidmaze tasks fall to near 0. On shorter tasks such as antmaze-medium and puzzle-3x3, the other two losses already do most of the work. The three variants give a clear order: no losses 33, all but the dynamics losses 47, all four losses 67.

AuthorsOpenReview ↗

(2/3) Response to Reviewer ccxi

W2. "The dynamics losses and latent unrolling is closely related to model-based RL while the approach is only compared against model-free GCRL approaches."

We added MAC (Park et al., 2025)[1], a recent offline model-based method evaluated on OGBench. MAC learns a dynamics model over action chunks and trains its value function on model rollouts. We implemented MAC in the same goal-conditioned setting as the other baselines: the same data, network sizes, DDPG+BC policy extraction, and evaluation as OGBench. These are our runs in this shared setting. Details are in App. E.5 (all tasks in Table 9):

EnvironmentGCTD3MACMSPR
antmaze-medium796899
antmaze-large253698
antmaze-giant0168
humanoidmaze-medium26667
humanoidmaze-large6058
Avg. (12)461467

The two methods use a dynamics model in different ways. MAC generates rollouts with its model and trains the critic on them, so the goal signal has to travel back along predicted trajectories. MSPR uses its dynamics heads, including the latent unroll, only to train the encoder. The agent never uses them to act or to compute value targets, so model errors do not enter the critic, and the distance to the goal comes directly from the Monte-Carlo return. The gap grows with the horizon (Figure 7): 68 vs. 99 on antmaze-medium, 1 vs. 68 on antmaze-giant. In this setting MAC is also low on manipulation (App. E.5, Table 9).

W3 / Q2. "Novelty unclear. The architecture and objectives are quite similar to MR.Q" / "What is the contribution of this paper?"

MSPR builds on MR.Q, and the paper says so. To test whether the MR.Q objectives are enough for goal-reaching, we trained the same GCTD3 agent (n=3n=3, same BC weights and discounts) on two other encoders: one trained with MR.Q's losses adapted to goals (GC-MR.Q), and the MSPR network with no predictive losses (No-aux) (selected tasks; all tasks in Table 8, App. E.4):

EnvironmentGCTD3No-auxGC-MR.QMSPR
antmaze-large250198
antmaze-giant00068
humanoidmaze-medium26242167
humanoidmaze-large60458
puzzle-4x488838793
Avg. (12)46333867
  • GC-MR.Q. GC-MR.Q reaches 38 on average. Here the reward predictor only estimates the single-step binary reward (whether the agent has reached the goal), so its latent space gets no direct signal of the distance to the goal. GC-MR.Q therefore fails on the long-horizon tasks, the same pattern as removing Lrew\mathcal L_{\mathrm{rew}}, while short-horizon manipulation stays close to MSPR.
  • No-aux. The MSPR network, all four loss weights set to 0, encoder trained by the critic's TD loss. No-aux reaches 33 on average, below the backbone (46). Its critic's value falls to the lower bound −1/(1−γ)-1/(1-\gamma) on both antmaze tasks in Figure 6 (App. E.4) within 75k steps, and does not recover. This shows that the gain comes from MSPR's predictive losses, not from the network.

How the analysis differs. MR.Q's latent-MDP results are for a single-task policy that reads only zsz_s; they ask the latent to predict the reward and the next state. A goal-conditioned policy reads the state and the goal together, and a latent can predict reward and next state well while (zs,zg)(z_s, z_g) still does not show the goal-directed action (Sec. 4, "The three terms are distinct"). Prop. 3 states this extra condition: if Lg-act\mathcal L_{g\text{-}act} is small, there is a policy on (zs,zg)(z_s, z_g) whose one-step gap to the optimal action is bounded. MR.Q's analysis has no such condition. Prop. 2 also differs: it uses the Monte-Carlo return, which carries the distance to the goal, instead of the one-step reward.

Our contribution is therefore:

  1. which predictions help a goal-conditioned encoder under sparse reward, with one loss and one upper bound for each (W1);
  2. controlled comparisons that place the gain in these losses, not in the network (No-aux), the MR.Q losses (GC-MR.Q), or model-based value learning (MAC);
  3. evidence that the same encoder raises the average of other backbones (GCIQL 45 to 56, GCBC 10 to 16; W4).

Reference

  1. Park, Kwanyoung, et al. "Scalable offline model-based RL with action chunks." International Conference on Learning Representations. Vol. 2026. (arXiv:2512.08108)
AuthorsOpenReview ↗

(3/3) Response to Reviewer ccxi

W4 / Q3. "Only TD3+BC as backbone is considered" / "Could the representation be used with e.g. CRL or GCIQL?"

Yes. We kept the encoder and its four losses unchanged and replaced GCTD3 with GCIQL or GCBC. Each agent keeps its own hyperparameters (selected tasks; all tasks in Table 7, App. E.3):

EnvironmentGCBC+ MSPRGCIQL+ MSPRGCTD3MSPR
antmaze-large244240862598
antmaze-giant010163068
humanoidmaze-medium82434592667
scene51035604245
pointmaze-medium91366437754
cube-single6660426365
Avg. (12)101645564667
  • GCIQL rises from 45 to 56, with the largest gains on long-horizon locomotion. It loses on pointmaze and cube, where its own networks are already strong; we report this in App. E.3.
  • GCBC rises from 10 to 16. GCBC has no critic and uses only zsz_s, so it cannot use the distance information that Lrew\mathcal L_{rew} puts into zsaz_{sa}, and a smaller gain is expected.
  • GCIQL and GCTD3 both learn from state and action, so they use the whole encoder (zsz_s and zsaz_{sa}), and that is where the encoder helps most.

Changes in the paper

  • Sec. 2: how MSPR differs from MR.Q (Monte-Carlo return vs. one-step reward) and from MAC (dynamics used only to train the encoder).
  • Sec. 4: each proposition is an upper bound, not a guarantee on the return; the meaning of "multi-scale".
  • Sec. 5 (RQ1): the role of each loss; No-dyn. Sec. 5 (RQ2): No-aux, GC-MR.Q, and other backbones.
  • App. E.3: MSPR with GCIQL and GCBC (Table 7).
  • App. E.4: No-aux, No-dyn, GC-MR.Q (Table 8), and No-aux training curves (Figure 6).
  • App. E.5: comparison with MAC (Table 9, Figure 7).
Reviewer_skW5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes what kind of representations are necessary in offline goal-conditioned reinforcement learning, as learning representations that enable high performance can be difficult in sparse reward scenarios. The authors argue that successful representation learning requires that immediate transition dynamics are captured and that goal-directed actions and successor states are predictable from their latent representations. Additionally, the representations must support learning the goal-conditioned returns efficiently. Based on these conditions, the authors propose MSPR, a method that trains a state/goal encoder with four prediction losses, and adds a goal-conditioned TD3-BC agent on top. Among the methods that MSPR is compared to, it has the best state-based average performance and the second-best performance on pixel tasks.

Critical strengths
  • The paper closely matches the hierarchical state-of-the-art method, HIQL, on the evaluated state-based navigation and play tasks without hierarchy, and outperforms the other flat methods on average, showing evidence that the proposed representation achieves high performance when combined with a goal-conditioned TD3-BC backbone. The method, MSPR, clearly outperforms the backbone GCTD3.
  • The ablations and experiments are thorough. For instance, the paper performs per-environment LOO ablations, clear comparisons to the backbone GCTD3, and stitch datasets. This makes it possible to see where the gains are coming from. Furthermore, the method is simple and cheap (flat, single policy).
Critical weaknesses
  • The paper makes the claim that dynamical alignment, behavioral alignment, and temporal alignment are necessary features for goal-aware representations. However, the empirical evaluation shows this through ablations on one single algorithm. The jump from the specific case of a proposed algorithm to a general claim that these components are demanded (L108-109, L472), could be better supported. Additionally, three of the four "multi-scale predictors" (L142) predict one step ahead and the "global" components just condition on the goal, which is confusing. The paper makes the claim that the temporal alignment "provides the critic with a pre-organized substrate," which could be defined more clearly and specifically evaluated. The entire gap to HIQL in state-based environments comes from 3 environments: the puzzle-play environments and cube-single-play. However, on these specific environments, the GCTD3 backbone performs comparably to MSPR (Table 3 in E.1), indicating that the gap to HIQL most likely comes from the selected backbone.
  • The paper is generally incremental, with the method very closely resembling MR.Q, with the additions being goal-conditioning, Monte-Carlo targets, and the two goal heads. These are arguably standard components. The goal-conditioned next-latent head is similar to HIQL's high-level policy converted into a loss, and with the next latent being the prediction target. Therefore, the paper's algorithmic novelty mostly consists of combining existing, well-known components, and the novelty depends on the quality of the analysis and empirical evaluation. The frame stacking is a confounding factor in the pixel-based tasks in the main results (Table 1). Among methods in Table 1, only MSPR uses frame stacking and without it the performance drops.
Questions for authors
  • Could you elaborate on what specifically makes the method "multi-scale"
  • Did you try HIQL and other baseline methods with the same frame stacking on the pixel tasks?
  • How compatible would the MSPR encoder be with a different backbone or another agent? Or how would some other representation perform on the GCTD3. This would help separate, what part of the contribution is dependent on the chosen backbone GCTD3.
  • What does "pre-organized substrate" mean?
5 comments
AuthorsOpenReview ↗

(1/3) Response to Reviewer skW5

We thank the reviewer for the detailed comments, and for noting the thorough ablations and that the method is simple. We answer each point below. Changes in the revised paper are shown in blue.

W1. "The paper makes the claim that dynamical alignment, behavioral alignment, and temporal alignment are necessary … the empirical evaluation shows this through ablations on one single algorithm."

We agree that "necessary" and "demands" were too strong. Propositions 1 to 3 bound the error that each input of the agent can cause. They do not bound the final return, and a method that lacks a condition does not have to fail. We changed the text:

  • Sec. 1: the three conditions "can be beneficial".
  • Sec. 4: "Each proposition is an upper bound on what the latent can support; none is a guarantee on the final return."
  • Sec. 2 and 6: "identifies predictive conditions that help in" and "benefits from representations that capture" (previously "required" and "demands").

The evidence is now wider than one algorithm. In Table 2, removing any single loss lowers both the state and the pixel average. We also tested the encoder with two more agents (GCIQL 45 to 56, GCBC 10 to 16), and other encoders on the same GCTD3 agent (Q3 below).

Q1 (and W1). "Could you elaborate on what specifically makes the method multi-scale?" / "three of the four multi-scale predictors predict one step ahead"

The reviewer is right that three heads predict one step. "Scale" refers to the time span of the information the target carries, not to the prediction step:

  • Ldyn\mathcal L_{dyn}: one step, unrolled in latent space over H=5H=5 steps.
  • Lg-dyn\mathcal L_{g\text{-}dyn}, Lg-act\mathcal L_{g\text{-}act}: one step, but conditioned on a goal that can be hundreds of steps ahead. To predict well, they must encode where the agent is heading over that span.
  • Lrew\mathcal L_{rew}: the Monte-Carlo return over the rest of the trajectory.

We added this definition to Sec. 4: "We call the representation multi-scale because the targets cover different time spans."

W2 / Q4. "pre-organized substrate" … "What does 'pre-organized substrate' mean?"

We removed the phrase. Sec. 1 now says that temporal alignment makes the latent reflect goal-conditioned returns, "so that the latent representation already holds an estimate of how far the goal is, and the critic does not have to find it from sparse rewards alone."

Concretely, frewf_{rew} predicts the Monte-Carlo return from (zsa,zg)(z_{sa}, z_g). With rewards in {−1,0}\{-1, 0\}, this return is set by how many steps remain until the goal. So states that are a similar number of steps from the goal get similar returns, and the latent holds an estimate of the distance to the goal. So the critic does not have to recover this distance from the sparse {−1,0}\{-1, 0\} reward alone. Prop. 2 makes this precise: the best error of Lrew\mathcal L_{rew} equals a noise term fixed by the data plus the smallest value error that any critic on the latent can reach, so lowering Lrew\mathcal L_{rew} lowers the value error the latent forces on the critic. We test this in two ways:

  1. Without Lrew\mathcal L_{rew}, the long-horizon tasks drop to 0 (antmaze-giant 68 to 0, humanoidmaze-large 58 to 0; Table 2).
  2. GC-MR.Q, whose encoder predicts only the one-step reward, fails on the same tasks (W3).
AuthorsOpenReview ↗

(2/3) Response to Reviewer skW5

W2. "The entire gap to HIQL in state-based environments comes from 3 environments … the gap to HIQL most likely comes from the selected backbone."

Our goal is a single method that performs well across both navigation and manipulation. To evaluate this, the benchmark spans regimes where each baseline family excels: long-horizon navigation, where hierarchical methods like HIQL are strongest, and manipulation, where flat value-based methods (GCIQL, DualGCIVL, and our GCTD3 backbone) already perform well. While we agree that GCTD3 is already competitive on manipulation and that MSPR improves upon an already strong backbone, the same backbone falls far behind HIQL on locomotion. On locomotion and navigation, the backbone falls far behind HIQL and our learned representation closes most of this gap without changing the underlying agent or its hyperparameters. Crucially, MSPR achieves this with minimal tuning: it uses a single BC weight per domain, the same discount rule as the baselines, and identical loss weights across all 17 environments (from Table 1):

State-basedGCTD3MSPRHIQL
Locomotion (7)367080
Manipulation (5)606317
All (12)466754

So both parts are needed: the backbone gives manipulation, and the representation gives locomotion. On the long horizon mazes, the representation lifts the backbone from 0 to 68 (antmaze-giant) and from 6 to 58 (humanoidmaze-large), matches with HIQL (Table 1). We no longer claim that MSPR beats HIQL. Sec. 5 now states the balance instead: MSPR's weaker domain (manipulation) averages 63, against 45 for the next best method.

W3. "The paper is generally incremental, with the method very closely resembling MR.Q … The goal-conditioned next-latent head is similar to HIQL's high-level policy."

MSPR builds on MR.Q, and the paper says so. To test whether the MR.Q objectives are enough for goal-reaching, we trained the same GCTD3 agent on other encoders (selected tasks, all tasks in Table 8, App. E.4):

EnvironmentGCTD3No-auxGC-MR.QNo-dynMSPR
antmaze-large25019098
antmaze-giant0003268
humanoidmaze-medium262421467
humanoidmaze-large604058
cube-single6353595265
Avg. (12)4633384767
  • GC-MR.Q reaches 38 on average. Here the reward predictor only estimates the single-step binary reward (whether the agent has reached the goal), so its latent space gets no direct signal of the distance to the goal. GC-MR.Q therefore fails on the long-horizon tasks, the same pattern as removing Lrew\mathcal L_{\mathrm{rew}}, while short-horizon manipulation stays close to MSPR.
  • No-aux Here MSPR network, all four loss weights set to 0, encoder trained by the critic's TD loss, No-aux reaches 33, below the backbone (46). Its critic's value falls to the lower bound −1/(1−γ)-1/(1-\gamma) on both antmaze tasks in Figure 6 (App. E.4) within 75k steps, and does not recover.
  • No-dyn (MSPR without Ldyn\mathcal L_{dyn} and Lg-dyn\mathcal L_{g\text{-}dyn}) reaches 47.

This helps us isolate the gain from the MSPR's predictive loss compared to No-aux (network alone) and the GC-MR.Q (MR.Q losses).

Relation to HIQL. Both Lg-dyn\mathcal L_{g\text{-}dyn} and HIQL's high-level policy use the same signal: the next state along a goal-directed trajectory. However, they use it differently. HIQL turns this signal into a second policy that generates subgoals at test time, requiring a tuned subgoal horizon per domain (10, 25, or 100 steps). In MSPR, Lg-dyn\mathcal L_{g\text{-}dyn} serves as a training signal for the encoder, keeping the agent flat with a single policy and no subgoal hyperparameter. MSPR is equal to HIQL on antmaze-giant (68) and humanoidmaze-large (58), and keeps the manipulation strength of a flat agent (63 vs. 17 for HIQL).

AuthorsOpenReview ↗

(3/3) Response to Reviewer skW5

Q3. "How compatible would the MSPR encoder be with a different backbone or another agent? Or how would some other representation perform on the GCTD3?"

We ran both halves of this question.

  • The MSPR encoder on other agents. We kept the encoder and its four losses unchanged and replaced GCTD3 with GCIQL or GCBC. The encoder raises GCIQL from 45 to 56 and GCBC from 10 to 16 (Table 7, App. E.3). The largest gains are on long-horizon locomotion (GCIQL on antmaze-giant: 1 to 63). GCIQL loses on pointmaze and cube, where its own networks are already strong.
  • Other representations on GCTD3. This is the W3 table. With the same agent, an encoder trained by the TD loss (No-aux) reaches 33, the MR.Q losses (GC-MR.Q) 38, MSPR without the dynamics losses (No-dyn) 47, and the full MSPR losses 67. The backbone with its own networks reaches 46.

In short, the gains come with the encoder when paired with different RL backbones, and ablations on the same backbone show that the MSPR losses produce this improvement.

W4 / Q2. "The frame stacking is a confounding factor in the pixel-based tasks" / "Did you try HIQL and other baseline methods with the same frame stacking?"

Yes. We re-ran all eight baselines with frame stacking (3 frames, 4 seeds; Table 4, App. E.1):

EnvironmentGCBCGCIVLGCIQLQRLCRLDual-GCIVLDual-CRLHIQLMSPR
visual-antmaze-medium24332704092819690
visual-antmaze-large912854767428381
visual-cube-single63574535567206682
visual-scene731364746315656
visual-puzzle-3x32320201130
Avg. with FS92329133054356068
Avg. single frame64224104449317555

Methods differ in whether frame stacking helps them. MSPR gains the most (55 to 68), and GCBC, GCIQL, QRL, and both Dual methods gain 3 to 5 points. GCIVL, CRL, and HIQL do better with single frames, in line with OGBench. With frame stacking for all methods, MSPR has the highest pixel average. With each method in its own best setting, HIQL leads (75, single frame), and we say so in the paper. Table 1 now reports each method in its best setting, and Table 4 shows both settings for every method.

Changes in the paper

  • Sec. 1, 2, 4, 6: the conditions are described as beneficial, not required; "pre-organized substrate" is replaced; the meaning of "multi-scale" is added (Sec. 4).
  • Sec. 5: the HIQL comparison is stated as balance across domains; Table 1 reports each pixel method in its best setting.
  • App. D: hyperparameters of all methods (Table 3).
  • App. E.1: all pixel methods with and without frame stacking (Table 4).
  • App. E.3, E.4: MSPR with GCIQL and GCBC (Table 7); No-aux, No-dyn, GC-MR.Q (Table 8, Figure 6).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.