acceptodds
Under review as a conference paper at ICLR 2027

InternW0-: An Embodied World Model Bridging Predictive Dynamics and Actions

20 views

Abstract

World Action Models (WAMs) have emerged as a promising paradigm for generalist robot manipulation by jointly modeling visual dynamics and action generation. A central challenge is how to effectively integrate complementary priors from large-scale pretrained models—including visual dynamics, scene semantics, and geometric and motion understanding—into a unified framework. To this end, we introduce InternW0-, a unified World Action Model that brings together pretrained visual dynamics, scene-level semantic understanding, 4D geometric and motion priors, and action generation within a directed Mixture-of-Transformers framework. Within the World–Action MoT, a pretrained video expert and an action expert interact under scene-grounded semantic guidance from a frozen VLM, while a pretrained 4D foundation model injects geometric and motion priors through training-only distillation. To further translate predictive visual dynamics into representations directly useful for action prediction, we introduce Causal Imprint, which learns future-relevant scene changes from training-only future supervision and makes these predictive representations directly available to the action expert without requiring future-video rollout at inference. To support large-scale joint training, we construct a heterogeneous corpus spanning robot demonstrations, UMI data, egocentric human demonstrations, and Ego2Robot data. We carefully curate and filter these diverse sources, resulting in over 20K hours of processed training data. We pretrain InternW0- on this heterogeneous corpus and demonstrate strong performance across diverse simulation benchmarks and real-robot platforms. We will open-source the model code and weights, infrastructure, and data-processing pipeline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Embodied Intelligence, Robot 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 3 reviews, above 85% of the others.
Reviewer_xUGSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents InternW0-Δ, a unified World Action Model that brings together pretrained visual dynamics, scene-level semantic understanding, 4D geometric and motion priors, and action generation within a directed Mixture-of-Transformers framework. Its key proposed component, Causal Imprint, uses future latent-difference supervision and future-feature alignment to learn predictive representations from observed context. The model is pretrained on heterogeneous robot and human data and evaluated on four simulation benchmarks and four real-robot platforms.

Critical strengths
  1. Strong empirical performance across diverse benchmarks. InternW0-Δ achieves 92.8% on LIBERO-Plus, 71.9% on RoboTwin Clean2Random, and 23.91% average success on RoboDojo, outperforming the reported WAM baselines.

  2. Substantial engineering and data contributions. The authors provide a detailed heterogeneous-data processing pipeline and an efficient inference implementation, reducing reported controller round-trip latency from 780.5 ms to 152.8 ms. These are meaningful practical contributions beyond the model architecture.

Critical weaknesses
  1. The methodological novelty of Causal Imprint is not sufficiently established. Future-supervised representations for action generation, training-time video supervision, and geometric feature distillation have close precedents in Fast-WAM, InternVLA-A1.5, Track4Action, and 4D-WAM. The proposed combination may be useful, but the current ablations do not isolate the distinctive value of predicting latent differences rather than directly supervising future representations. In Table 5, CI alone improves success from 69.08% to 70.80%, while future-feature alignment yields the larger subsequent gain to 76.45%. The evidence therefore supports the complete predictive-supervision pipeline more strongly than the specific claim that change-oriented imprinting is the key innovation.
Questions for authors
  1. Can the authors provide a matched ablation comparing latent-difference supervision against direct future-latent supervision, while keeping CI tokens, attention, and future-feature alignment unchanged?
2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

1. Isolating latent-difference supervision and feature alignment

Both latent-difference supervision and feature alignment are applied to the Causal Imprint tokens. We evaluated each objective independently, as well as their combination, under matched training settings.

ModelSpatialObjectGoalLongOverall
Feature alignment only72.15% (1733/2402)77.20% (1944/2518)45.97% (1191/2591)64.47% (1624/2519)64.73% (6492/10030)
Latent-difference supervision only79.18% (1902/2402)83.64% (2106/2518)55.00% (1425/2591)66.22% (1668/2519)70.80% (7101/10030)
Full81.85% (1996/2402)89.20% (2246/2518)60.67% (1572/2591)74.79% (1884/2519)76.45% (7668/10030)

Latent-difference supervision alone improves the overall success rate from 69.08% to 70.80%, whereas feature alignment alone decreases it to 64.73%. In contrast, combining the two objectives yields an overall success rate of 76.45%, improving over the baseline by 76.45% - 69.08% = 7.37% and over latent-difference supervision alone by 76.45% - 70.80% = 5.65%.

These results suggest that the two objectives play complementary roles. In particular, the gains produced by feature alignment in the full model should not be interpreted as evidence that feature alignment alone accounts for the overall improvement; rather, its effectiveness emerges when it is combined with latent-difference supervision.

2. Matched comparison with direct future-latent supervision

To directly examine the value of change-oriented imprinting, we replaced latent-difference supervision with direct future-latent supervision while keeping the remaining settings unchanged. This comparison isolates the choice of predictive target within the Causal Imprint framework.

ModelSpatialObjectGoalLongOverall
Future-latent supervision81.35% (1954/2402)86.42% (2176/2518)63.87% (1655/2591)69.59% (1753/2519)75.15% (7538/10030)
Latent-difference supervision81.85% (1996/2402)89.20% (2246/2518)60.67% (1572/2591)74.79% (1884/2519)76.45% (7668/10030)

Direct future-latent supervision achieves an overall success rate of 75.15%, compared with 76.45% for latent-difference supervision. Thus, replacing latent differences with direct future-latent prediction reduces overall performance by 76.45% - 75.15% = 1.30%.

Thank you for the additional ablations. However, the reported 1.30% advantage over direct future-latent supervision (75.15% vs 76.45%) is modest and varies across task groups. I therefore encourage the authors to temper the corresponding claims and incorporate these comparisons into the revised manuscript. Considering the overall empirical and engineering contributions, I maintain my Weak Acceptance rating.

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

The paper presents InternW0-∆, an embodied world-action model that combines video and vision-language representations with predictive latent supervision and 4D distillation. It is pretrained on heterogeneous embodied data and evaluated on four simulation benchmarks and several robot platforms, with an inference design that avoids generating future videos.

Critical strengths
  1. The evaluation covers robustness, long-horizon manipulation, and deployment across different robot platforms. Strong simulation results, together with real-world demonstrations, provide meaningful evidence of the system’s practical value.
  2. The paper addresses deployment alongside model training. Its inference optimizations and detailed descriptions of data processing and training make the system contribution more substantial than the individual architectural changes alone.
Critical weaknesses
  1. The distinct value of Causal Imprint is not sufficiently isolated. In Table 5, latent-difference supervision adds 1.72 points, followed by another 5.65 points from feature alignment. These are conditional gains in a cumulative ablation. Without alignment-only and direct future-latent prediction controls, it remains unclear whether predicting differences offers an advantage over simpler predictive supervision.
  2. The component ablations do not cover the final pretraining regime. Table 5 trains variants from scratch and reaches 78.37%, whereas the pretrained model achieves 92.8%. The reported component gains therefore do not establish their contributions to the final model. Similarly, Tables 16–17 test individual pretraining sources but do not isolate Ego2Robot’s benefit within the full mixture. Training-seed variability is also unreported for the main results.
Questions for authors
  1. Under matched training settings, how does CI compare with alignment-only supervision and direct prediction of future latents?
  2. Do the gains from CI, alignment, and 4D distillation persist in the final pretrained model? Please provide controlled ablations that distinguish pretraining contributions from post-training effects.
2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

1. Feature-alignment-only and latent-difference-only controls

Both latent-difference supervision and feature alignment are applied to the Causal Imprint tokens. We additionally evaluated each supervision objective independently under matched training settings. The results are summarized below.

ModelSpatialObjectGoalLongOverall
Feature alignment only72.15% (1733/2402)77.20% (1944/2518)45.97% (1191/2591)64.47% (1624/2519)64.73% (6492/10030)
Latent-difference supervision only79.18% (1902/2402)83.64% (2106/2518)55.00% (1425/2591)66.22% (1668/2519)70.80% (7101/10030)
Full81.85% (1996/2402)89.20% (2246/2518)60.67% (1572/2591)74.79% (1884/2519)76.45% (7668/10030)

Relative to the baseline without either supervision objective, which achieves an overall success rate of 69.08%, latent-difference supervision alone improves performance to 70.80%, corresponding to a gain of 70.80% - 69.08% = 1.72%. In contrast, feature alignment alone reduces the overall success rate to 64.73%. However, jointly applying both objectives improves the overall success rate to 76.45%, representing gains of 76.45% - 69.08% = 7.37% over the baseline and 76.45% - 70.80% = 5.65% over latent-difference supervision alone.

These results suggest a complementary relationship between the two supervision objectives. Feature alignment is most effective when combined with latent-difference supervision, rather than when used as a standalone objective, supporting our design choice of jointly applying both forms of supervision to the Causal Imprint tokens.

2. Comparison with direct future-latent supervision

To isolate the effect of predicting latent differences, we conducted an additional experiment in which latent-difference supervision was replaced with direct future-latent supervision, while keeping the remaining experimental settings unchanged. The results are as follows.

ModelSpatialObjectGoalLongOverall
Future-latent supervision81.35% (1954/2402)86.42% (2176/2518)63.87% (1655/2591)69.59% (1753/2519)75.15% (7538/10030)
Latent-difference supervision81.85% (1996/2402)89.20% (2246/2518)60.67% (1572/2591)74.79% (1884/2519)76.45% (7668/10030)

Replacing latent-difference supervision with direct future-latent supervision reduces the overall success rate on LIBERO-Plus from 76.45% to 75.15%, corresponding to a decrease of 76.45% - 75.15% = 1.30%.

3. Component contributions in the final pretrained regime

Given the scale of our pretraining dataset and the substantial computational and time costs associated with large-scale pretraining, repeating the pretraining procedure for every architectural variant is prohibitively expensive. Therefore, we adopt a consistent and computationally tractable model-development protocol: we validate our design choices through controlled experiments during post-training, rather than repeatedly conducting large-scale pretraining for each variant. The final model is subsequently obtained through large-scale pretraining followed by post-training on LIBERO.

Importantly, our evaluation of these design choices is not limited to in-distribution performance. All variants are evaluated on LIBERO-Plus, an out-of-distribution (OOD) benchmark, under the same experimental protocol. This OOD evaluation provides a stringent test of whether the proposed components improve generalization to unseen conditions, rather than merely improving performance on the training distribution. We therefore believe that these controlled post-training comparisons, together with the OOD evaluation, provide meaningful evidence for the effectiveness of our model designs while keeping the experimental cost tractable.

4. Training-seed variability

To assess training-seed variability, we started from the same final pretrained checkpoint reported in the paper and conducted three independent post-training runs on LIBERO using different random seeds. The resulting overall success rates on LIBERO-Plus were 92.8%, 92.7%, and 92.3%, respectively.

All three runs are reported here. No additional runs were conducted and omitted, and these results were not selected from a larger pool of experiments. The narrow spread of 92.8%−92.3%=0.592.8\% - 92.3\% = 0.5 percentage points between the highest and lowest results suggests that the reported performance is not sensitive to the choice of post-training seed and can be reproduced consistently.

Thank you for the reply. The additional controls for Causal Imprint and the reported seed variability adequately address my main concerns. I maintain my original score.

Reviewer_A5NLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes InternW0-Δ, a world-action model that combines pretrained video dynamics, scene understanding, and geometric representations for robot control. The main technical contributions include Causal Imprint, which learns future visual changes without requiring video generation at inference, and 4D-aware representation distillation. The model is pretrained on over 20K hours of heterogeneous demonstrations. Experiments on LIBERO-Plus, RoboTwin2.0, EBench, and RoboDojo show strong performance, with additional demonstrations on four real-robot platforms.

Critical strengths

The paper evaluates four simulation benchmarks and multiple real-robot platforms. InternW0-Δ achieves 92.8% on LIBERO-Plus and 71.9% on RoboTwin Clean2Random, outperforming the compared baselines. The results across different benchmarks are encouraging.

Causal Imprint uses future information during training while avoiding expensive video generation at inference. The ablations show that combining future-feature alignment, semantic guidance, and 4D distillation improves performance. The extensive data processing and deployment infrastructure are also valuable contributions.

Critical weaknesses

Table 5 shows that adding the pretrained VLM provides a large improvement, while Causal Imprint and 4D distillation bring smaller additional gains. Also, the final ablation reaches 78.37% on LIBERO-Plus, compared with 92.8% for the full model in Table 1. It is unclear how much of the final improvement comes from the proposed methods versus large-scale pretraining and the pretrained backbones.

The paper demonstrates 11 tasks across four robot platforms, but reports quantitative success rates for only two tasks in the pretraining comparison. There are also no direct comparisons with competing policies on these real robots. This makes it difficult to judge the practical improvement of the proposed method.

Questions for authors

What explains the difference between the 78.37% result in Table 5 and the 92.8% result in Table 1? Can the authors provide ablations of Causal Imprint and 4D distillation using the full pretrained model?

Can the authors report quantitative success rates across all 11 real-world tasks, ideally with comparisons against existing VLA or WAM baselines?

1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

1. Difference between the 78.37% result in Table 5 and the 92.8% result in Table 1

The two results were obtained under different training regimes. Specifically, the 78.37% result in Table 5 was obtained by initializing the video expert from Wan2.2-5B TI2V and randomly initializing the action expert, followed by training on LIBERO and evaluation on LIBERO-Plus, without the large-scale pretraining stage. In contrast, the 92.8% result in Table 1 was obtained by first pretraining the model with the same initialization for 245k steps, subsequently post-training the resulting checkpoint on LIBERO, and finally evaluating it on LIBERO-Plus. Therefore, the performance difference primarily arises from the additional large-scale pretraining stage rather than from inconsistent evaluation protocols.

Given the scale of our pretraining dataset and the substantial computational and time costs associated with large-scale pretraining, repeating the pretraining procedure for ablation study is prohibitively expensive. Therefore, we adopt a consistent and computationally tractable model-development protocol: we validate our design choices through controlled experiments during post-training, rather than repeatedly conducting large-scale pretraining for each variant. The final model is subsequently obtained through large-scale pretraining followed by post-training on LIBERO.

Importantly, our evaluation of these design choices is not limited to in-distribution performance. All variants are evaluated on LIBERO-Plus, an out-of-distribution (OOD) benchmark, under the same experimental protocol. This OOD evaluation provides a stringent test of whether the proposed components improve generalization to unseen conditions, rather than merely improving performance on the training distribution. We therefore believe that these controlled post-training comparisons, together with the OOD evaluation, provide meaningful evidence for the effectiveness of our model designs while keeping the experimental cost tractable.

2. Additional real-robot evaluation

We additionally evaluated our method on the RoboDojo real-robot benchmark, which provides direct comparisons with existing VLA and WAM policies across three robot platforms.

As shown below, our method, InternW0-Δ, ranks first among the compared methods and achieves an average success rate of 31.70%. In comparison, OpenWAM-α\alpha achieves an average success rate of 24.40%, corresponding to an absolute improvement of 7.30 percentage points for our method. These results provide further evidence of the practical effectiveness of our approach in real-world robotic settings.

ModelTypeAverageARX X5PiperPiper X
StarVLAVLA4.10 / 1.70%0.00 / 0.00%12.00 / 5.00%0.50 / 0.00%
Pi-0VLA5.80 / 1.70%8.10 / 0.00%9.00 / 5.00%0.30 / 0.00%
GR00T-N1.7VLA5.90 / 1.70%9.30 / 1.70%6.80 / 3.30%1.40 / 0.00%
X-VLAVLA7.60 / 3.30%6.80 / 1.70%15.90 / 8.30%0.10 / 0.00%
Xiaomi-Robotics-0VLA7.90 / 3.90%15.00 / 8.30%7.60 / 3.30%1.20 / 0.00%
GalaxeaVLAVLA9.00 / 4.40%3.40 / 0.00%22.00 / 13.30%1.70 / 0.00%
InternVLA-A1VLA12.00 / 7.20%10.00 / 3.30%23.20 / 18.30%2.70 / 0.00%
Pi-05VLA22.90 / 12.80%27.70 / 13.30%32.00 / 21.70%9.10 / 3.30%
OpenWAM-α\alphaWAM37.60 / 24.40%39.00 / 23.30%46.70 / 36.70%27.20 / 13.33%
InternW0-ΔWAM38.90 / 31.70%59.70 / 50.00%39.30 / 31.70%17.80 / 13.33%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.