acceptodds
Under review as a conference paper at ICLR 2027

RoboCTR: Cross-Modal Trajectory Rewarding for World Model-Based Robotic Planning

Abstract

Action-conditioned world models require a planning score that evaluates predicted trajectories with respect to the task goal. We propose RoboCTR, a Cross-Modal Trajectory Reward (CTR) interface that scores a latent rollout against the available text, goal-image, and demonstration-video goals. Stage I learns a goal-conditioned per-step critic; Stage II freezes this critic and uses its gradients to shape the V-JEPA2 predictor; receding-horizon MPC then reuses the same critic to rank candidate action sequences. We further introduce RoboTrain, an eight-task simulation collection with keyframe-aligned subtask instructions and positive and negative trajectory segments. On seven RoboTwin2 tasks, RoboCTR achieves 84.3% macro success versus 79.7% for 0, a gain of 4.6 percentage points, with the largest gains on Pick Dual Bottles and Place Object in Basket. Additional ablations examine the training stages, reward interface, goal modalities, task decomposition, and guidance weight. We also present qualitative demonstrations on a physical manipulator.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI

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_eHWrOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to use a world model for robot planning. A world model can predict the outcome of candidate actions, but the robot still needs a way to judge which predicted outcome is closer to the goal. RoboCTR addresses this by training a reward model that scores predicted trajectories using the available goal information, such as text instructions, goal images, or demonstration videos.

The authors first train this reward model on robot demonstrations, then use it to guide the world-model predictor and to rank candidate action sequences during MPC. They also introduce RoboTrain, an eight-task simulation dataset. On seven RoboTwin2 Easy tasks, RoboCTR reports 84.3% average success versus 79.7% for π0, and provides qualitative demonstrations on an ALOHA dual-arm platform.

Critical strengths
  1. S1. The shared reward interface is coherent and intuitive. Using one goal-conditioned critic both to shape the world-model predictor and to rank MPC candidates provides a clean connection between training and planning. Supporting text, goal images, and demonstration videos within the same interface is also practically appealing.

  2. S2. The paper reports encouraging simulation results and broad design studies. RoboCTR improves over π0 on four of the seven reported tasks, including gains of 15 and 22 percentage points on Pick Dual Bottles and Place Object Basket. The experiments also cover training stages, reward interfaces, goal modalities, task decomposition, and planning/guidance weights.

Critical weaknesses

W1. The progress score may not reflect real progress toward success. Equation 5 assigns higher scores to later frames in a successful window, but being later in time does not necessarily mean the robot is closer to completing the task. This leaves a gap between the learned score and actual task progress.

Questions for authors
  1. Table 5 arithmetic. The Default row reports 84.3%, but the deltas are computed against 82.7%; for example, 74.1% is listed as −8.6 instead of −10.2. Please correct the table.
Reviewer_JeyoOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces RoboCTR, a Cross-Modal Tra-jectory Reward interface that scores a latent rollout against multimudal goals.

Critical strengths

The paper provides clear experimental results and ablations. Analyses of the planning horizon, the number of candidates, and the loss weights provide useful insights.

Critical weaknesses
  • The experimental evaluation is limited. The real-world experiments only consist of two simple manipulation tasks. RoboTwin 2.0 contains 50 tasks, but the paper evaluates only seven in the Easy setting, without explaining the selection criteria. Broader evaluation would help establish how well the method generalizes beyond this subset.
  • Selecting actions by evaluating multiple rollouts in a latent world model is an established approach, as demonstrated by DINO-WM, V-JEPA 2, and LeWM. More discussion and direct comparisons would help establish the additional value of the proposed shared reward interface.
  • This paper introduce a method to learn multi-modal reward for latent planning, but it is not clear why it is necessary and what advantages it offers.
Questions for authors
  • Could the authors clarify why combining multiple goal modalities improves performance?
  • What is the planning latency under the default setting of 512 candidates?
1 comment
Area_Chair_MmbiOpenReview ↗

Reminder to respond to the authors’ rebuttal

With one day left in the rebuttal period, please read the authors’ response and leave a comment. It would help the authors to know which of your concerns they have addressed and which questions remain open, so they can make the most of the time left.

Thanks for your time. Your constructive feedback helps the authors improve their work and strengthens the ICLR community.

Best regards,

Area Chair

Reviewer_ka4VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RoboCTR uses a learned critic to score latent trajectories against text, image, or demonstration-video goals. Stage I trains the critic with progress regression, instruction contrast, and a monotonicity loss. Stage II freezes the critic and trains the V-JEPA2 action-conditioned predictor with prediction loss and critic-output regression. MPC then uses the same critic to rank candidate trajectories and executes the first action. The paper introduces RoboTrain, with 240 simulation episodes across eight tasks, and reports 84.3% macro success on seven RoboTwin2 Easy tasks versus 79.7% for pi0. Physical-robot results are qualitative.

Critical strengths
  1. CTR has a useful role in both training and planning. It guides predictor adaptation and then scores candidates during MPC. The equations allow gradients through the frozen critic, and valid failure transitions still train the predictor even though they receive no progress targets. The guidance-weight ablation is a useful control because the zero-guidance condition retains reconstruction training.

  2. The main benchmark shows worthwhile gains. Table 1 uses 100 rollouts per task. Compared with pi0, RoboCTR gains 15 percentage points on Pick Dual Bottles and 22 on Place Object Basket, with a 4.6-point improvement in the seven-task macro average. It also loses ground on some tasks, which helps show where the method is useful.

Critical weaknesses
  1. The progress labels admit a simple shortcut. Equation 5 gives every eligible four-step window the same targets, approximately [0.500, 0.574, 0.719, 1.000]. Instruction contrast uses mismatches from other tasks, while failure, stall, and backtracking windows receive no progress or trend targets (Appendices A–B). A critic that recognizes the task and tracks position within a window could fit these objectives without reliably distinguishing progress from stalling on the same task. The appendix also acknowledges the limits of instruction contrast for learning semantic progress.

  2. The evidence for same-task ranking is incomplete. The improvement from 74.3% to 82.7% in Table 8(a) shows that CTR guidance helps. The 52.0% success with wrong goals in Figure 6(b) shows that goal information matters, but does not tell us whether CTR ranks competing trajectories for one goal correctly. AUROC and ECE could be informative here, yet their labels, sample construction, and calibration procedure are not specified. I cannot tell what aspect of scoring these metrics validate.

  3. The baseline inputs are not clear enough to assess the comparison. RoboCTR receives goal references and ordered subgoals, and uses completion predicates that are defined only abstractly. Table 1 does not make clear which goal inputs and task-specific supervision the baselines receive. These details matter for attributing the gains to CTR.

  4. The default results need to be reconciled. Table 5 lists 84.3%, while its text and reported differences use 82.7%. For instance, 74.1% is 8.6 points below 82.7%. Tables 3 and 8(b) also report different results for the ordered configuration without enough detail to identify the experimental difference. The ablations need a clear mapping of tasks, rollout counts, seeds, and settings to the main benchmark.

Questions for authors
  1. How well does CTR rank successful, stalled, and backtracking trajectories for the same task, compared with a baseline that uses task recognition and position within the window? Please give the labels and held-out splits for AUROC/ECE, and report scoring performance separately on encoded observations and predicted trajectories.

  2. Could using the same CTR in Stage II and MPC amplify discrepancies between predicted scores and actual outcomes? Please compare guidance-enabled training with the reconstruction-only control (lambda_CTR = 0) on held-out trajectories. Alongside latent prediction error, how well do predicted CTR scores agree with independently measured progress or success? The concern is whether guidance improves prediction or mainly makes predictions conform to the critic's biases.

  3. For Table 1, what training data, goal modalities, subgoal annotations, and completion information does each method receive? What are the implemented completion predicates, and do they use simulator information beyond the stated observations and proprioception?

  4. Which task sets, rollout counts, seeds, and configurations produced the 82.7% and 84.3% defaults? Please reconcile these results, correct Table 5, and report variation across runs for the main comparison.

1 comment
Area_Chair_MmbiOpenReview ↗

Reminder to respond to the authors’ rebuttal

With one day left in the rebuttal period, please read the authors’ response and leave a comment. It would help the authors to know which of your concerns they have addressed and which questions remain open, so they can make the most of the time left.

Thanks for your time. Your constructive feedback helps the authors improve their work and strengthens the ICLR community.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.