PAPER DEEP DIVE

Paper世界模型World Model

Action from Adjacent Set in Physical Space Outperforms the Best Prediction in World Models

Controllers based on sampling and latent world models assign a predicted terminal cost to each candidate action sequence, choose the minimum, execute its first action block, and replan. This rule can fail even when the terminal cost perfectly and accurately reflects the true task objective in the physical world. Residual prediction error can give an infeasible sequence an anomalously low cost, and a larger proposal pool gives such errors more chances to outrank feasible alternatives. We call this conditional failure proposal overgeneration. In Cube candidate execution audits, increasing the total proposal budget from 72 to 288 reduces the feasibility of selection by minimum latent cost from .375 to .062 for position targets and from .344 to .031 for targets defined by position and yaw, although every larger pool contains a feasible sequence. We introduce Adjacent Set Action Reconstruction (ASAR). Among proposals with low cost, ASAR measures density from standardized early action prefixes and reconstructs a full sequence from an adjacent set with a light anchor from the sequence with minimum cost. On a Carry and Release evaluation set of 75 queries, Kernel ASAR improves event completion success over matching selection by 28.0, 24.0, and 18.7 percentage points under latent cost and by 18.7, 20.0, and 17.3 points under a trajectory reachability cost at 72, 144, and 288 proposals. Analysis of finite proposal pools characterizes selection risk from the lower tail, separation by a related radius support statistic, and sequence containment under an explicit local feasibility condition.

Liangyu Li, Qingwen Liu, Mingqing LiuJuly 26, 20264 min read

中文

Action from Adjacent Set in Physical Space Outperforms the Best Prediction in World Models

Authors: Liangyu Li, Qingwen Liu, Mingqing Liu | Institution: Tongji University | arXiv: 2607.23602v1

One-Sentence Summary

This paper identifies "proposal overgeneration" in sampling-based latent world model controllers — larger candidate pools reduce minimum-cost selection reliability — and proposes ASAR (Adjacent Set Action Reconstruction): identifying an adjacent set among low-cost candidates via prefix distances and reconstructing a new action sequence via weighted aggregation. On Carry-and-Release, ASAR improves over minimum-cost selection by 28.0/24.0/18.7 percentage points at 72/144/288 proposals.

Background and Motivation

Sampling-based latent world model controllers assign predicted terminal costs to candidate action sequences, select the minimum, execute its first action block, and replan. This can fail even when the terminal cost perfectly reflects the true objective: residual prediction error can give an infeasible sequence an anomalously low cost, and larger pools give such errors more chances to outrank feasible alternatives.

Figure 1: Minimum-cost sequence (red) is isolated in prefix space; closed-loop execution leaves cube 21.2cm from target. 12 low-cost sequences (green) form an adjacent set; Kernel ASAR reconstruction (blue) leaves cube 3.8cm away.

Proposal overgeneration: Increasing budget from 72 to 288 reduces minimum latent cost selection feasibility from .375 to .062 (position targets) and .344 to .031 (position+yaw), despite every larger pool containing feasible sequences.

Problem Formulation

Encoder $e$ maps observations to latent space. Candidate sequence $\mathbf{A}_i \in \mathcal{A}^H$. Predicted terminal:

$$\hat{\mathbf{z}}_{i,H}=\hat{f}_{H}(\mathbf{z}_{t},\mathbf{A}_{i})$$

Terminal cost $q_i = Q(\hat{\mathbf{z}}_{i,H}, \mathbf{z}_g)$. Feasible set:

$$\mathcal{E}_{\epsilon}(\mathbf{x}_{t},o_{g})=\{\mathbf{A}\in\mathcal{A}^{H}:\ell_{H}(F_{H}(\mathbf{x}_{t},\mathbf{A}),o_{g})\leq\epsilon\}$$

Ranking blocker — infeasible proposal $b$ with $q_{b}<\min_{i:Y_{i}=1}q_{i}$.

Proposal Scaling under Model Error

Top-$k$ feasibility decomposition:

$$\Pr(T_{B,k})=\Pr(P_{B})\,\Pr(T_{B,k}\mid P_{B})$$

Probability of low-cost infeasible sequence in finite pool:

$$\Pr\!\left(\min_{b:Y_{b}=0}q_{b}\leq t\right)=1-\left(1-F_{-}(t)\right)^{n_{-}}$$

Target	Cost	B=72 Top-20	B=288 Top-20	B=72 Blockers	B=288 Blockers
Position	Latent	.969	.656	4.31	16.91
Position	Reachability	.875	.750	6.66	16.66
Pos+yaw	Latent	.969	.625	4.34	18.31
Pos+yaw	Reachability	.906	.719	7.59	17.75

Figure 2: Top-20 feasibility declines from 72 to 288 proposals; mean blockers increase. Reachability cost changes levels but doesn't remove the effect.

ASAR Method

Adjacent Set Identification

Among $M$ lowest-cost candidates $\mathcal{R}_M(Q)$, vectorize first $h$ action blocks into prefix $\mathbf{p}_i$. Local prefix isolation score:

$$u_{i}=\frac{1}{K}\sum_{j\in\mathcal{N}_{K}(i)}\|\mathbf{p}_{i}-\mathbf{p}_{j}\|_{2}$$

Hoeffding bound on radius support reversal: if $\mathbb{E}S_r(e) \geq p_+$ (feasible) and $\mathbb{E}S_r(b) \leq p_-$ (infeasible blocker):

$$\Pr(S_{r}(e)\leq S_{r}(b))\leq 2\exp\!\left[-\frac{n(p_{+}-p_{-})^{2}}{2}\right]$$

Full Action Sequence Reconstruction

Kernel ASAR selects candidate $c$ with smallest $u_i$, identifies adjacent set $\mathcal{S}$, assigns weights:

$$w_{i}\propto\exp\!\left(-\frac{\|\mathbf{p}_{i}-\mathbf{p}_{c}\|_{2}^{2}}{\tau_{d}}-\lambda\tilde{q}_{i}\right)$$

Reconstruction:

$$\mathbf{A}_{\mathrm{rec}}=(1-\alpha)\sum_{i\in\mathcal{S}}w_{i}\mathbf{A}_{i}+\alpha\mathbf{A}_{Q}$$

Conditional containment: $\|\mathbf{A}_{\mathrm{rec}}-\mathbf{A}_{c}\|\leq(1-\alpha)R+\alpha D$ where $R$ is neighborhood radius and $D=\|\mathbf{A}_Q - \mathbf{A}_c\|$.

flowchart TD
    A["CEM generates candidate pool
B action sequences"] --> B["World model predicts terminal
ẑ_i,H = f̂_H(z_t, A_i)"]
    B --> C["Terminal cost ranking
q_i = Q(ẑ_i,H, z_g)"]
    C --> D["Keep M lowest-cost
R_M(Q)"]
    D --> E["Prefix isolation score
u_i = KNN mean distance"]
    E --> F["Select candidate c with min u_i
Identify adjacent set S"]
    F --> G["Kernel-weighted reconstruction
A_rec = (1-α)Σw_i·A_i + α·A_Q"]
    G --> H["Execute A_rec first action block
Replan"]

Experimental Results

Carry-and-Release (75 queries, 44 seeds)

Proposals	Latent-min	Latent-least iso.	Latent-ASAR	Reach-min	Reach-ASAR
72	.067	.187	.347	.147	.333
144	.053	.213	.293	.133	.333
288	.133	.147	.320	.133	.307

ASAR gains 28.0/24.0/18.7 pp under latent cost (95% CIs exclude 0); 18.7/20.0/17.3 pp under reachability cost.

Reconstruction Beats Existing Candidate Selection

73-query targeted test: least isolated existing sequence .493, Kernel ASAR reconstruction .630 (+13.7 pp, CI [1.9, 23.8]). Independent world model weights: ASAR .738 vs .717.

Where ASAR Helps

At 288 proposals: ASAR rescues 19 queries, loses 5. On 42 long multi-behavior queries (grasp-lift-transport-lower-release), ASAR .048 < baseline .119 — single adjacent set cannot supply multiple distinct behaviors.

Limitations

Limited effectiveness on long-horizon queries requiring multiple distinct behaviors.
Conditional containment is local; doesn't extend to arbitrary pools.
Only validated on Cube manipulation; scalability to complex tasks unverified.
Prefix isolation is a density proxy, not a feasibility label.

Conclusion and Outlook

This paper identifies proposal overgeneration in world model control — larger candidate pools reduce minimum-cost selection reliability under residual model error. ASAR identifies adjacent sets in prefix space among low-cost candidates and reconstructs new action sequences via weighted aggregation, without modifying the world model, proposal generator, or terminal cost. The reconstructed sequence need not exist in the sampled pool, providing intermediate behaviors unavailable through existing candidate selection. Formal analysis provides guarantees on lower-tail ranking errors, local support separation, and sequence containment.

Key insight: "Action from adjacent set in physical space outperforms the best prediction in world models" — when residual prediction error makes minimum-cost selection unreliable, leveraging prefix distance relations within the candidate pool to identify locally regular regions and reconstruct actions is a more direct solution than improving predictors or increasing pool size.

Deep analysis generated by RobotWorld paper-detail-generator based on full-text reading | arXiv:2607.23602v1

Paper 世界模型 World Model

SOURCE LINKS

https://arxiv.org/abs/2607.23602

Related Papers

Pre-training Visual Dexterity in Simulation

Large-scale pre-training has made robot policy fine-tuning increasingly data-efficient, but this progress has largely been driven by datasets and embodiments built around simple parallel-jaw grippers. Dexterous, multi-fingered hands remain comparatively data-starved because real teleoperation is costly to scale, while human hand video is off-embodiment and requires lossy pose estimation and retargeting. We introduce Simulation Pre-training for Dexterity (SPD), a pre-training framework for dexterous manipulation that uses data entirely collected in simulation. In SPD, humans manipulate virtual objects inside a VR headset, enabling on-embodiment trajectories and robot-free collection. With the help of five operators, we collect 75 hours of multi-task dexterous manipulation over one week, and use it to pre-train a causal transformer on a sequence modeling objective. We study the benefits of simulation pre-training on real-world tasks by fine-tuning on 1-2 hours of physical demonstrations on a 56-DoF bimanual dexterous setup. We find that our approach outperforms training behavior cloning policies from scratch, showing that simulation teleoperation is a viable pre-training source for real-world dexterous manipulation. We perform ablation studies, measuring the benefits of history conditioning and short action chunks for reactive control.

灵巧操作灵巧手预训练Aug 16, 2026

FA-RDP: A Frequency-Adaptive Reactive Diffusion Policy for Contact-Rich Manipulation

In contact-rich manipulation, action multimodality and reactivity dominate different stages of a single episode. Before contact, multiple trajectories might be equally valid, making it important to preserve diverse action modes. After contact, geometric constraints and force limits narrow the solution space, while successful execution demands rapid responses to force feedback. However, standard diffusion policies use a fixed inference frequency and sampling steps throughout the episode, forcing a fundamental compromise: low-frequency, multi-step sampling better preserves pre-contact multimodality but responds slowly to force feedback, whereas high-frequency sampling improves reactivity but tends to collapse distinct pre-contact modes. To resolve this tradeoff, we present FA-RDP, a frequency-adaptive reactive diffusion policy. A shared multi-frequency visual-force Transformer predicts action chunks at both low and high frequencies, while a learned multimodality indicator dynamically selects multi-step low-frequency sampling before contact and one-step high-frequency sampling as action ambiguity decreases. We further introduce Manifold Consistency Distillation (MCD), which reparameterizes the diffusion network to predict actions on the robot action manifold while retaining DDPM-based residual supervision. Experiments on three contact-rich manipulation tasks show that FA-RDP achieves the highest success rate while preserving diverse pre-contact trajectory modes. Code and videos are available at https://fa-rdp.github.io.

PaperJul 30, 2026

PhiZero: A World Model Built Around Physical Language

We introduce PhiZero, a physical world model built around physical language, a compact discrete representation of world-state transitions. Existing physical world models typically predict future videos directly in pixel space, leaving the underlying world dynamics implicit within high-dimensional visual predictors. Motivated by humans' ability to abstract predictive structure from visual experience and organize it in natural language for explicit reasoning, we learn physical language from in-the-wild videos through self-supervision and use it to explicitly reason about how the physical world evolves. Accordingly, PhiZero adopts a reason-then-render paradigm: it first infers future world evolution as a physical-language sequence and then renders the inferred transitions into videos. Extensive experiments across generation and understanding benchmarks validate the ability of PhiZero to model physically coherent world evolution. We further show its potential for realistic and interactive world modeling, fine-grained action-conditioned simulation, and zero-shot motion transfer.

PaperJul 30, 2026

Booster Lab: A Data-Centric Pipeline for Learning Deployable Humanoid Locomotion Policies

Humanoid robot motion learning requires not only task-oriented control policies but also physically feasible and natural behaviors that can be transferred to real robots. However, robot-feasible motion data are often scarce: raw human demonstrations may be incompatible with the robot morphology, open-source clips vary in quality, and simulation-collected robot trajectories still require feasibility checking. To address these challenges, we propose a data-centric training and deployment pipeline that integrates motion data curation, real-to-sim model adaptation, AMP-based reinforcement learning, and sim-to-real deployment. We validate the framework on the Booster T1 robot and further provide preliminary cross-platform validation on Booster K1.

步态优化人形机器人AMPJun 26, 2026

[●●●]FETCHING

←Back to Papers

PAPER DEEP DIVE

Paper世界模型World Model

Action from Adjacent Set in Physical Space Outperforms the Best Prediction in World Models

Liangyu Li, Qingwen Liu, Mingqing LiuJuly 26, 20264 min read

中文

Action from Adjacent Set in Physical Space Outperforms the Best Prediction in World Models

Authors: Liangyu Li, Qingwen Liu, Mingqing Liu | Institution: Tongji University | arXiv: 2607.23602v1

One-Sentence Summary

Background and Motivation

Problem Formulation

Encoder $e$ maps observations to latent space. Candidate sequence $\mathbf{A}_i \in \mathcal{A}^H$. Predicted terminal:

$$\hat{\mathbf{z}}_{i,H}=\hat{f}_{H}(\mathbf{z}_{t},\mathbf{A}_{i})$$

Terminal cost $q_i = Q(\hat{\mathbf{z}}_{i,H}, \mathbf{z}_g)$. Feasible set:

$$\mathcal{E}_{\epsilon}(\mathbf{x}_{t},o_{g})=\{\mathbf{A}\in\mathcal{A}^{H}:\ell_{H}(F_{H}(\mathbf{x}_{t},\mathbf{A}),o_{g})\leq\epsilon\}$$

Ranking blocker — infeasible proposal $b$ with $q_{b}<\min_{i:Y_{i}=1}q_{i}$.

Proposal Scaling under Model Error

Top-$k$ feasibility decomposition:

$$\Pr(T_{B,k})=\Pr(P_{B})\,\Pr(T_{B,k}\mid P_{B})$$

Probability of low-cost infeasible sequence in finite pool:

$$\Pr\!\left(\min_{b:Y_{b}=0}q_{b}\leq t\right)=1-\left(1-F_{-}(t)\right)^{n_{-}}$$

Target	Cost	B=72 Top-20	B=288 Top-20	B=72 Blockers	B=288 Blockers
Position	Latent	.969	.656	4.31	16.91
Position	Reachability	.875	.750	6.66	16.66
Pos+yaw	Latent	.969	.625	4.34	18.31
Pos+yaw	Reachability	.906	.719	7.59	17.75

Figure 2: Top-20 feasibility declines from 72 to 288 proposals; mean blockers increase. Reachability cost changes levels but doesn't remove the effect.

ASAR Method

Adjacent Set Identification

Among $M$ lowest-cost candidates $\mathcal{R}_M(Q)$, vectorize first $h$ action blocks into prefix $\mathbf{p}_i$. Local prefix isolation score:

$$u_{i}=\frac{1}{K}\sum_{j\in\mathcal{N}_{K}(i)}\|\mathbf{p}_{i}-\mathbf{p}_{j}\|_{2}$$

Hoeffding bound on radius support reversal: if $\mathbb{E}S_r(e) \geq p_+$ (feasible) and $\mathbb{E}S_r(b) \leq p_-$ (infeasible blocker):

$$\Pr(S_{r}(e)\leq S_{r}(b))\leq 2\exp\!\left[-\frac{n(p_{+}-p_{-})^{2}}{2}\right]$$

Full Action Sequence Reconstruction

Kernel ASAR selects candidate $c$ with smallest $u_i$, identifies adjacent set $\mathcal{S}$, assigns weights:

$$w_{i}\propto\exp\!\left(-\frac{\|\mathbf{p}_{i}-\mathbf{p}_{c}\|_{2}^{2}}{\tau_{d}}-\lambda\tilde{q}_{i}\right)$$

Reconstruction:

$$\mathbf{A}_{\mathrm{rec}}=(1-\alpha)\sum_{i\in\mathcal{S}}w_{i}\mathbf{A}_{i}+\alpha\mathbf{A}_{Q}$$

Conditional containment: $\|\mathbf{A}_{\mathrm{rec}}-\mathbf{A}_{c}\|\leq(1-\alpha)R+\alpha D$ where $R$ is neighborhood radius and $D=\|\mathbf{A}_Q - \mathbf{A}_c\|$.

flowchart TD
    A["CEM generates candidate pool
B action sequences"] --> B["World model predicts terminal
ẑ_i,H = f̂_H(z_t, A_i)"]
    B --> C["Terminal cost ranking
q_i = Q(ẑ_i,H, z_g)"]
    C --> D["Keep M lowest-cost
R_M(Q)"]
    D --> E["Prefix isolation score
u_i = KNN mean distance"]
    E --> F["Select candidate c with min u_i
Identify adjacent set S"]
    F --> G["Kernel-weighted reconstruction
A_rec = (1-α)Σw_i·A_i + α·A_Q"]
    G --> H["Execute A_rec first action block
Replan"]

Experimental Results

Carry-and-Release (75 queries, 44 seeds)

Proposals	Latent-min	Latent-least iso.	Latent-ASAR	Reach-min	Reach-ASAR
72	.067	.187	.347	.147	.333
144	.053	.213	.293	.133	.333
288	.133	.147	.320	.133	.307

ASAR gains 28.0/24.0/18.7 pp under latent cost (95% CIs exclude 0); 18.7/20.0/17.3 pp under reachability cost.

Reconstruction Beats Existing Candidate Selection

73-query targeted test: least isolated existing sequence .493, Kernel ASAR reconstruction .630 (+13.7 pp, CI [1.9, 23.8]). Independent world model weights: ASAR .738 vs .717.

Where ASAR Helps

Limitations

Limited effectiveness on long-horizon queries requiring multiple distinct behaviors.
Conditional containment is local; doesn't extend to arbitrary pools.
Only validated on Cube manipulation; scalability to complex tasks unverified.
Prefix isolation is a density proxy, not a feasibility label.

Conclusion and Outlook

Deep analysis generated by RobotWorld paper-detail-generator based on full-text reading | arXiv:2607.23602v1

Paper 世界模型 World Model

SOURCE LINKS

https://arxiv.org/abs/2607.23602