Robotics · arXiv:2609.27656 v1 · submitted 23 Sep 2026 · 23 references
InternW0: A Foundational Physical World Model for Efficient Real-World Interactions
The paper says
“Physical intelligence requires more than predicting how the world may evolve: predictions must remain actionable as the world continues to change.”
more of the abstract
We introduce InternW0, the first instantiation of the InternW physical world model series from Shanghai AI Laboratory, built around omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling under partial observations and external influences. InternW0 jointly learns future visual dynamics and continuous robot control through an asymmetric video–action architecture with flow matching. A high-capacity video expert provides longer-horizon predictive context, while a lightweight action expert operates at a faster timescale. Instead of regenerating the future for every action update, InternW0 reuses layerwise K/V and adapts it to newly observed states through observation-conditioned context routing. Domain-specific interfaces and soft prompts support heterogeneous embodiments, while contact-aware post-training incorporates force and tactile signals for contact-rich manipulation. We train InternW0 on approximately 7,200 hours of heterogeneous robot and egocentric data, including EgoLab, a 275-hour real-laboratory egocentric dataset. Evaluation spans simulation benchmarks and real-world scientific tasks, including a 15-stage metal–organic framework synthesis workflow and 5-stage contact- and force-aware dexterous manipulation for general-purpose quantitative pipetting. These results advance scalable, asynchronous, and science-native physical world models for universal and efficient real-world interactions.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 24 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A serious, well-engineered WAM report with genuinely strong end-to-end numbers, but the headline efficiency claim rests on a K/V caching editor that is never ablated against fresh video regeneration. This line is wrong
4 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 23 references say.
The paper says, at §2.1, Eqs. (2)–(4); §6.1
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The central efficiency claim (16.47 Hz, 60.73 ms, 3.13× over Fast-WAM in Figure 7) measures only the action critical path; the paper itself states that video-plan generation is excluded from this measurement (§6.1). Consequently, the accuracy side of the speed/accuracy tradeoff rests entirely on the observation-conditioned chunk K/V editor, which is never ablated.
more of objection 1
Although the zero-initialized residual in Eq. (4) guarantees the editor starts as an identity mapping, no experiment shows that its learned corrections preserve downstream success at the level of regenerating the video context per action chunk, or that they beat naive reuse of the unedited cache. Without such an ablation, the statement that asynchronous execution allows reuse 'without removing [prediction's] influence on control' (Section 6, opening) is unverified, and the 3.13× speedup could come at a real but unmeasured accuracy cost. I recommend an ablation under the same post-training and evaluation protocol with three settings: (i) full video regeneration per action chunk, (ii) cached context edited by the K/V editor, and (iii) cached context used without editing.
The paper says, at §5.1, Tables 5–7; §4.2.1
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
No error bars, confidence intervals, or significance tests appear anywhere in the manuscript. Real-robot results use 15 trials per task, so a single trial moves an episode-level success rate by about 6.7 percentage points; several reported differences are within this granularity (e.g., Make Sandwich 73.3 vs. 73.3 and Sort Tubes 88.9 vs. 86.7 in Table 5).
more of objection 2
Likewise, the claim in §4.2.1 that a 0.16 pp clean/randomized gap 'shows that the model maintains highly stable performance' is not supported without variance estimates, and the LIBERO margin of 0.1 pp over LingBot-VA (Table 2) is within the noise of 50 rollouts per task. I request that the authors report means with standard deviations or confidence intervals and flag which inter-method differences are statistically distinguishable.
The paper says, at §4.2.2, Table 4
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The headline Clean2Random margin (75.60% vs. 67.3% for GigaBrain-0.7, an 8.3 pp advantage) rests on baseline numbers cited from other groups' papers rather than rerun under a shared protocol. The randomized-split margin is only 0.1 pp (68.0 vs. 67.9), while the clean-split margin is 16.4 pp (83.2 vs. 66.8); if GigaBrain-0.7 and the other baselines used different post-training schedules, rollout counts, or task subsets, the average margin may not transfer.
more of objection 3
The paper should state, for each baseline in Tables 2–4, whether the number comes from the authors' own replication (with rollout counts and hardware) or from the original publication, and should temper claims of superiority by the margin's split structure.
The paper says, at §4.2.1, Table 3
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
On RoboTwin 2.0-Full, InternW0 (93.12%) is below ABot-M0.5 (94.10%) and OpenWAM-α (93.60%), so the presentation should be candid that InternW0 is not state-of-the-art in this in-domain setting. The text's emphasis on the small clean/randomized gap as evidence of stability is not warranted without error bars, and it should be made explicit that the Full setting is a competitive-but-not-leading result.
more of objection 4
This matters because the Clean2Random result is the stronger claim and readers need to see that the Full setting does not contradict it through evaluation-protocol differences.
6 minor objections
The paper says, at §2.1, flow-matching paragraph
Pith says: minor objection 1 of 6
The sentence 'Flow times are sampled using the shifted continuous flow-matching schedule... independent flow times for individual action chunks' is immediately followed by a near-duplicate sentence ('Flow times are sampled per video example and per action chunk.'); one of the two should be removed.
The paper says, at Table 3, header row
Pith says: minor objection 2 of 6
The method entry 'π0.5π0.5' appears to be a formatting error for π0.5; the duplicate token should be corrected.
show 4 more
The paper says, at Figure 7
Pith says: minor objection 3 of 6
The caption line 'One action-generation update · update rate = 1 / latency' is confusing because the two panels use different quantities (latency in ms and update rate in Hz). Clarify that the 60.73 ms figure excludes asynchronous video-plan generation, as stated in §6.1, and label the axes of both panels consistently.
The paper says, at §2.1, capacity descriptors
Pith says: minor objection 4 of 6
The paper repeatedly calls the action expert 'lightweight' and the video expert 'high-capacity' but never reports parameter counts, FLOPs, or latency of the two experts separately. Reporting these numbers is necessary to substantiate the capacity asymmetry that the asynchronous claim depends on.
The paper says, at §3, EgoLab
Pith says: minor objection 5 of 6
The EgoLab dataset records human participants with head-mounted cameras and wristband controls, but the manuscript includes no ethics, consent, or data-protection statement for these recordings; standard practice for human egocentric datasets is to include such a statement.
The paper says, at Throughout
Pith says: minor objection 6 of 6
Several typographical and reference issues: 'MoF Experimentsreproduces' is missing a space (§5.1); 'built on the top of Ray' is ungrammatical (§2.2); the GR00T-N1.7 reference has a malformed author list ('Nikita Cherniadev Fernando Castañeda'); and the bibliographic entry for the world-action-models survey has a journal title appended to an arXiv item. These should be cleaned up.
What would settle it
Re-run RoboTwin 2.0-Clean2Random (or LIBERO) with the chunk K/V editor disabled and the full video context regenerated from current observations before every 16-step action chunk, keeping the same pretrained checkpoint, post-training recipe, rollout counts, and evaluation protocol as Tables 3–4. Compare clean and randomized success rates, and also measure critical-path latency in this configuration. If success drops by more than roughly 2 percentage points on either split, the cached-editor design exchanges accuracy for speed and the central efficiency claim is not supported; if success is statistically indistinguishable, the efficiency/accuracy tradeoff is validated.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 24 Sep 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
What to know: this is a solid engineering contribution, not a flimsy one. InternW0 couples a video DiT (slow prediction) with a lightweight action DiT (fast control) through an observation-conditioned chunk K/V editor, and the end-to-end results are competitive: 98.6% LIBERO, 75.60% average on RoboTwin C2R, a new 275-hour egocentric lab dataset (EgoLab), and real wet-lab demos (MoF synthesis at 68.4% progress, quantitative pipetting at 65.3%) at a 16.47 Hz policy update rate. The attribution is honest: context routing is credited to AHA-WAM, soft prompts to X-VLA, the backbone style to the MoT line, flow matching to pi0. Provenance is not being hidden. The architecture is concrete (Eqs. 1–6), the data pipeline shows real care with validity masks and split hygiene, and the subtask breakdowns in Tables 6–7 are the right way to report contact-rich results.
The stress-test concern is real, and I agree it is the decisive one: the K/V editor (Section 2.1, Eqs. 2–4) is never compared against fresh regeneration of the video context. The residual is zero-initialized, so the editor starts as an identity map and learns corrections, but no experiment shows those corrections hold accuracy at the level of full regeneration. And the paper itself states the latency number excludes video-plan generation (Section 6.1). So the 3.13x speedup and the accuracy margins cannot yet be jointly attributed to the architecture; the speed could be coming with an accuracy cost that is simply never measured. That is the one ablation I would require before trusting the efficiency story.
Secondary items, in proportion: baselines are cited from other groups' papers, not rerun under a shared protocol; there are no error bars anywhere; real-robot trials are 15 per task, so one trial moves a point by roughly 6–7; the EgoLab ablation runs on ~50h of video, which shows egocentric pretraining helps in a limited-data regime but does not validate the full-corpus scaling. None of these are fatal — the system works end-to-end as reported. What is unverified is the marginal contribution of the specific mechanism that the paper's novelty mostly hangs on. I also note the RoboTwin C2R margin over GigaBrain-0.7 comes almost entirely from the clean split; on randomized scenes the gap is 0.1 points.
For whom: anyone working on world-action models, asynchronous prediction/control, or cross-embodiment pretraining. EgoLab itself is a real resource worth having.
Recommendation: send it to serious peer review. The K/V editor ablation should be required in revision; reproduced baselines and error bars, strongly requested. This is a real contribution deserving referee time, not a desk reject.
The strongest objection
The load-bearing efficiency claim is that the observation-conditioned chunk K/V editor (Section 2.1, Eqs. 2–4) preserves future-video prediction quality while keeping video regeneration off the critical action path, enabling 16.47 Hz (Fig 7, Section 6.1) without hurting success. The paper provides no ablation comparing this cached-editor configuration against fresh regeneration of the video context for each action chunk under the same evaluation protocol. Equation 4’s residual is zero-initialized, so the editor starts as an identity map and learns corrections, but no experiment shows those corrections keep downstream success at the level of full regeneration. The reported critical-path latency (60.73 ms) explicitly excludes video-plan generation, as the paper states: 'Video-plan generation is scheduled asynchronously and is therefore not included in this critical-path measurement' (Section 6.1). Thus the latency number measures only the fast action path, while the accuracy side of the tradeoff rests entirely on the unablated editor. Without a regeneration baseline, the headline margins (e.g., Clean2Random 75.60% average vs. 67.3% for GigaBrain-0.7) and the 3.13x speedup cannot be jointly attributed to the architecture; the speed gain could be coming at a real accuracy cost that the paper never measures. The baseline-reproduction concern is real but secondary; the editor ablation is the more decisive missing experiment.
The reader’s weakest_assumption identifies two concerns: unreproduced external baselines and the missing K/V editor ablation. I agree with both, but the missing editor ablation is the more load-bearing for the paper’s central architectural claim. The paper’s headline contribution is the asynchronous, duplex design that reuses cached predictive K/V through the observation-conditioned editor; if that editor degrades task success relative to fresh regeneration, then the reported speed advantage is achieved at an unmeasured accuracy cost, and the Clean2Random and real-world margins cannot be confidently attributed to the claimed mechanism. The reader’s verdict of CONDITIONAL already captures this, so I do not move the verdict. I set agreement_with_reader to 'partial' because the reader lists the editor ablation as one of several concerns rather than the primary one, and I would make it the decisive test for the efficiency claim. The concrete test I propose is a direct A/B comparison that would settle whether the cached-editor speedup preserves accuracy; this is feasible with the same benchmark and evaluation protocol already used in the paper.
The referee's report
If the joint speed–accuracy claim holds, the contribution is significant: an asynchronous duplex world-action architecture that decouples expensive future-video regeneration from fast closed-loop control while retaining observation-conditioned context adaptation, a large heterogeneous pretraining corpus with a new egocentric laboratory dataset, and real-robot demonstration on contact-rich scientific tasks (MOF synthesis, pipetting). The manuscript is strong on concreteness: the architecture is specified in equations (2)–(7), the data-recipe and cleaning rules are detailed enough to reproduce, and the internal arithmetic of the reported aggregates is consistent. The main load-bearing weakness is that the efficiency result rests on an unablated component: the paper never compares the cached K/V editor against fresh regeneration of the video context, so the speed gain and the accuracy numbers cannot yet be jointly attributed to the architecture. All benchmark and real-robot numbers also lack variance estimates, and the headline margins over cited baselines are not backed by a shared rerun protocol.
Pith's reading of the paper (a machine's summary, not the authors')
This paper presents InternW0, a physical world model trained to predict future video frames and generate continuous robot actions at once, then run on two separate timescales. The central claim is that a robot need not regenerate its imagined future on every control step: a fast action expert reuses cached predictive context from a slower video expert and adapts it to each new observation, keeping expensive video prediction off the latency-critical action path. Built on a mixture-of-transformers backbone with flow matching and contact- and force-aware post-training, InternW0 reports 98.6% success on LIBERO, 75.60% average on RoboTwin 2.0-Clean2Random, and progress rates of 68.4% and 65.3% on real-world metal-organic-framework synthesis and quantitative pipetting, while reaching a 16.47 Hz critical-path update rate, 3.13x the Fast-WAM baseline. The reason this matters is that a world model for physical interaction must deliver predictions soon enough to guide ongoing action, not merely predict accurately.
What carries the argument. The chunk K/V editor is the load-bearing mechanism. After the video expert generates a future plan, its layerwise keys $K_\ell$ and values $V_\ell$ are cached. For each action chunk $n$, the current RGB observation is encoded by a frozen DINOv3 encoder and summarized into routing queries $q_n$; the editor retrieves information from the cached context via attention, routes it back to the original video-token positions, and a zero-initialized projection predicts residual corrections $\Delta K_{\ell,n}$, $\Delta V_{\ell,n}$ gated by a learned scalar $\sigma(\gamma_\ell)$. The edited context $\tilde K_{\ell,n} = K_\ell + \sigma(\gamma_\ell)\Delta K_{\ell,n}$ is what the action expert consumes. Because the residual starts at zero, the editor begins as identity and learns to introduce observation-dependent corrections, allowing the system to reuse expensive predictive computation while remaining sensitive to divergence between the predicted future and the state that actually unfolds.
Core claim. The paper's central claim is that future-video prediction and robot control need not be synchronized. A high-capacity video expert predicts longer-horizon future frames on a slow schedule, while a lightweight action expert generates action chunks on a fast schedule and reads the video expert's cached layerwise key and value representations through an observation-conditioned editor. The editor retrieves task-relevant predictive information using the latest observed image as queries, routes it back to the original video-token positions, and predicts small residual corrections to the cached keys and values before the action expert consumes them, starting as an identity mapping through a zero-initialized projection. The paper argues this lets execution deviate from the imagined future without paying for a full regeneration, and that gradients flowing from the action objective into the shared video expert shape predictive representations for action usefulness rather than visual reconstruction alone. On that basis InternW0 reports 98.6% on LIBERO, 93.12% on RoboTwin 2.0-Full, 75.60% average on Clean2Random, 68.4% progress on a 15-stage metal-organic-framework synthesis, 65.3% progress on quantitative pipetting, and a 16.47 Hz critical-path update rate, 3.13x over Fast-WAM.
The premise it rests on. The load-bearing premise is that editing cached predictive context through the K/V editor keeps task success as high as regenerating a fresh future prediction would, a premise the paper asserts but does not test directly, and that the headline margins over cited baselines hold under a shared protocol the authors did not run.
If the paper is right (Pith's inferences, not the authors')
If reusable predictive context preserves control quality, world-model policies can decouple prediction and control timescales, removing video-generation latency from the closed loop without eliminating predictive influence on action.
Because gradients from the action objective flow through the editor into the shared video expert, predictive representations are shaped for action usefulness, implying that prediction quality and control quality co-improve during joint flow-matching training.
The limited-data ablation showing an 8.24-percentage-point gain in randomized Clean2Random success from adding 50 hours of EgoLab video to 50 hours of robot data implies that action-free human egocentric video contributes to out-of-distribution generalization.
Domain-specific soft prompts paired with a unified 37-dimensional action interface let one pretrained backbone post-train onto single-arm, bimanual, and dexterous-hand platforms without retraining the shared video expert.
Contact-aware post-training with joint action-wrench prediction gives force and tactile history a direct conditioning pathway into action chunks, which the paper links to the dexterous pipetting improvements.
If the editor truly preserves prediction quality at a fraction of the cost, the same cached-context routing idea transfers to any latent that is expensive to regenerate each step, such as 3D scene representations or long-horizon task plans, not only future video.
The paper does not ablate the editor against fresh regeneration, so the natural decisive experiment is editing cached context versus regenerating it at equal wall-clock budget, a test that would reveal any hidden success-rate cost of reuse.
Reusing cached context trades per-step predictive freshness for a shorter observation-to-action interval, so the right regime for the strategy depends on task dynamics; contact-rich tasks where an early wrong action damages the object may favor fresh regeneration despite its latency.
The reported margins are against other groups' published baseline numbers under different schedules and hardware, so a shared-protocol rerun would be needed to confirm that the 8.30-point Clean2Random lead over GigaBrain-0.7 transfers.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Physical intelligence requires more than predicting how the world may evolve: predictions must remain actionable as the world continues to change. We introduce InternW0, the first instantiation of the InternW physical world model series from Shanghai AI Laboratory, built around omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling under partial observations and external influences. InternW0 jointly learns future visual dynamics and continuous robot control through an asymmetric video–action architecture with flow matching. A high-capacity video expert provides longer-horizon predictive context, while a lightweight action expert operates at a faster timescale. Instead of regenerating the future for every action update, InternW0 reuses layerwise K/V and adapts it to newly observed states through observation-conditioned context routing. Domain-specific interfaces and soft prompts support heterogeneous embodiments, while contact-aware post-training incorporates force and tactile signals for contact-rich manipulation. We train InternW0 on approximately 7,200 hours of heterogeneous robot and egocentric data, including EgoLab, a 275-hour real-laboratory egocentric dataset. Evaluation spans simulation benchmarks and real-world scientific tasks, including a 15-stage metal–organic framework synthesis workflow and 5-stage contact- and force-aware dexterous manipulation for general-purpose quantitative pipetting. These results advance scalable, asynchronous, and science-native physical world models for universal and efficient real-world interactions.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-24
The authors' main claim, as the classifier picked it out 1.00
“These results advance scalable, asynchronous, and science-native physical world models for universal and efficient real-world interactions.”
How the main claim is stated: asserted 0.99
What the abstract itself states, as probabilities: claims first 0.86 · states limitation 0.08 · human participants 0.10 · controlled comparison 0.06 · randomized assignment 0.01 · states quantitative result 0.78 · compares with prior results 0.09 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Physical intelligence requires more than predicting how the world may evolve: predictions must remain actionable as the world continues to change. | background 0.87 | 0.07 | 0.11 | asserted 1.00 |
| We introduce InternW0, the first instantiation of the InternW physical world model series from Shanghai AI Laboratory, built around omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling under partial observations and external influences. | contribution 0.99 | 0.90 | 0.07 | asserted 1.00 |
| InternW0 jointly learns future visual dynamics and continuous robot control through an asymmetric video--action architecture with flow matching. | method 0.89 | 0.91 | 0.06 | asserted 1.00 |
| A high-capacity video expert provides longer-horizon predictive context, while a lightweight action expert operates at a faster timescale. | method 0.98 | 0.64 | 0.03 | asserted 1.00 |
| Instead of regenerating the future for every action update, InternW0 reuses layerwise K/V and adapts it to newly observed states through observation-conditioned context routing. | method 0.97 | 0.90 | 0.05 | asserted 1.00 |
| Domain-specific interfaces and soft prompts support heterogeneous embodiments, while contact-aware post-training incorporates force and tactile signals for contact-rich manipulation. | method 0.86 | 0.85 | 0.04 | asserted 1.00 |
| We train InternW0 on approximately 7,200 hours of heterogeneous robot and egocentric data, including EgoLab, a 275-hour real-laboratory egocentric dataset. | method 0.99 | 0.94 | 0.03 | asserted 1.00 |
| Evaluation spans simulation benchmarks and real-world scientific tasks, including a 15-stage metal--organic framework synthesis workflow and 5-stage contact- and force-aware dexterous manipulation for general-purpose quantitative pipetting. | result 0.46 | 0.94 | 0.09 | asserted 0.88 |
| These results advance scalable, asynchronous, and science-native physical world models for universal and efficient real-world interactions. | interpretation 1.00 | 0.46 | 0.83 | asserted 0.99 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.86
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-24: 1 request, 6,995 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (7)
Figures are the authors', reproduced from the paper.
References (23)
- [1] Lei Bai, Jiaqi Cao, Chiyu Chen, Guanzhou Chen, Kai Chen, Guangran Cheng, Erfei Cui, Xuanlang Dai, Shengyuan Ding, Shangheng Du, et al. Intern-S2-Preview: Scientific agentic foundation model.arXiv preprint arXiv:2608.13505, on PitharXiv
- [3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fu- sai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, on PithPith's reviewarXiv
- [9] doi: 10.1038/s41586-025-08744-2. URL https://www.nature.com/articles/s41586-025-08744-2. Nicklas Hansen, Hao Su, and Xiaolong Wang. TD-MPC2: Scalable, robust world models for con- tinuous control. InThe Twelfth International Conference on Learning Representations, on Pithdoi:10.1038/s41586-025-08744-2
- [10] URL https://arxiv.org/abs/2310.16828. InkStone, Shanghai AI Laboratory. InkStone: Ai research workspace.https://discovery-home. intern-ai.org.cn/home/. Accessed: September 22, on PithPith's reviewarXiv
- [11] Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al.π 0.5: a vision-language-action model with open-world generalization.arXiv preprint arXiv:2504.16054, on PithPith's reviewarXiv
- [13] URLhttps://arxiv.org/abs/2509.00576. Leslie Pack Kaelbling, Michael L. Littman, and Anthony R. Cassandra. Planning and acting in partially observable stochastic domains.Artificial Intelligence, 101(1–2):99–134, on PithPith's reviewarXiv
- [15] URLhttps://arxiv.org/abs/2508.07917. Michelle A. Lee, Yuke Zhu, Krishnan Srinivasan, Parth Shah, Silvio Savarese, Li Fei-Fei, Animesh Garg, and Jeannette Bohg. Making sense of vision and touch: Self-supervised learning of mul- timodal representations for contact-rich tasks. In... on PithPith's reviewarXiv
- [17] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864, on PithPith's reviewarXiv
- [19] GigaBrain Team, Angen Ye, Axiang Sun, Can Jin, Chenxi Cheng, Chong Shi, Dengke Shang, Dingqian Zhang, Guan Huang, Guangqiang Wang, et al. Gigabrain-0.7: Scaling embodied foundation models to emergent capabilities with a three-system architecture.arXiv preprint arXiv:2608.15875, on PithPith's reviewarXiv
- [20] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, on PithPith's reviewarXiv
- [21] Siyin Wang, Junhao Shi, Zhaoyang Fu, Xinzhe He, Feihong Liu, Chenchen Yang, Yikang Zhou, Zhaoye Fei, Jingjing Gong, Jinlan Fu, et al. World action models: The next frontier in embodied ai.arXiv preprint arXiv:2605.12090, 2026a. Yuran Wang, Siqiao Huang, Mingleyang Li, Chenhao ... on PithPith's reviewarXiv
- [22] Lishan Yang, Wenxuan Song, Xi Wang, Pingyue Sheng, Zheng Fang, Ziyang Zhou, Junjie He, Haodong Yan, Jiayi Chen, Nan Sun, et al. 4d-wam: Infusing spatiotemporal awareness into world action models through trajectory fields.arXiv preprint arXiv:2608.08023, 2026a. Yandan Yang, Shu... on PithPith's reviewarXiv
- [23] URLhttps://arxiv.org/abs/2602.15922. Jiawen Yu, Hairuo Liu, Qiaojun Yu, Jieji Ren, Ce Hao, Haitong Ding, Guangyu Huang, Guofan Huang, Yan Song, Panpan Cai, et al. Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation.Advances in Neural Information... on PithPith's reviewarXiv
- [25] URLhttps://arxiv. org/abs/2603.16666. Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolandos Alexandros Potamias. Hawor: World-space hand motion reconstruction from egocentric videos. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1805–1815. IEEE, on PithPith's reviewarXiv
- [26] Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yinan Zheng, Jiayin Zou, Yilun Chen, Jia Zeng, et al. X-vla: Soft-prompted transformer as scalable cross- embodiment vision-language-action model. InInternational Conference on Learning Representa... on Pith
- [1981] doi: 10.1115/1.3139652. URLhttps: //doi.org/10.1115/1.3139652. Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning.Proceedings of the IEEE, 109(5):612–634, on Pithdoi:10.1115/1.3139652
- [1998] doi: 10.1016/S0004-3702(98)00023-X. URLhttps://www.sciencedirect.com/science/article/pii/ S000437029800023X. Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Opti- mizing speed and success.arXiv preprint arXiv:2502.19645, on PithPith's reviewarXiv
- [2018] URL https://arxiv.org/abs/1803.10122. Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. InProceedings of the 36th Interna- tional Conference on Machine Learning, volume ... on PithPith's reviewarXiv
- [2019] URLhttps://arxiv.org/abs/1810.10191. Fuhao Li, Wenxuan Song, Han Zhao, Jingbo Wang, Pengxiang Ding, Donglin Wang, Long Zeng, and Haoang Li. Spatial forcing: Implicit spatial representation alignment for vision-language- action model. InInternational Conference on Learning Repr... on PithPith's reviewarXiv
- [2023] URLhttps://proceedings.mlr.press/ v205/chen23d.html. Jun Guo, Qiwei Li, Peiyan Li, Zilong Chen, Nan Sun, Yifei Su, Heyun Wang, Yuan Zhang, Xinghang Li, and Huaping Liu. Unified 4d world action modeling from video priors with asynchronous denoising.arXiv preprint arXiv:2604.26694, on PithPith's reviewarXiv
show 3 more
- [2024] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems.arXiv preprint arXiv:2503.06669, on PithPith's reviewarXiv
- [2025] Jisong Cai, Long Ling, Shiwei Chu, Zhongshan Liu, Jiayue Kang, Zhixuan Liang, Wenjie Xu, Yinan Mao, Weinan Zhang, Xiaokang Yang, et al. Aha-wam: Asynchronous horizon-adaptive world- action modeling with observation-guided context routing.arXiv preprint arXiv:2606.09811, 2026a.... on PithPith's reviewarXiv
- [2026] doi: 10.48550/arXiv.2608.13505. URLhttps:// arxiv.org/abs/2608.13505. Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chendong Xiang, Yinze Rong, et al. Motus: A unified latent action world model. In Proceedings of the IEEE/... on Pithdoi:10.48550/arxiv.2608.13505
Cited by 3 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- InfCoiL: Coordinated Planner-Controller Learning for Closed-Loop Physics-Based Human-Object Interaction cs.CV · Nov 2026
- Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models cs.CV · Nov 2026
- Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy cs.RO · Oct 2026
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment