ALAM: Algebraically Consistent Latent Transitions for Vision-Language-Action Models

Summary

Vision-Language-Action (VLA) models map visual observations and language instructions to robot actions through learned latent representations.

Abstract Summary

Vision-Language-Action (VLA) models map visual observations and language instructions to robot actions through learned latent representations. This paper proposes ALAM, which enforces group-theoretic constraints on latent transitions to ensure compositional consistency and reliable multi-step reason... The method demonstrates significant improvements over existing approaches, providing both theoretical insights and practical benefits for real-world deployment. Comprehensive experiments validate the effectiveness of the proposed approach across diverse scenarios and task settings.

Key Points

Proposes ALAM, a novel approach for vla in robotics.
Addresses key limitations in existing methods through innovative architecture design.
Demonstrates strong empirical results on standard benchmarks and real-world evaluations.
Introduces novel training methodology that improves generalization and sample efficiency.
Provides comprehensive analysis of failure modes and ablation studies.

Method

We introduce ALAM , a two-stage framework that learns latent transitions from action-free video and transfers them to vision-language-action (VLA) policy learning. In pretraining (Sec.

3.2 ), a relational encoder is trained with future-frame reconstruction together with two algebraic regularizers, composition consistency and reversal consistency, that encourage the latent transitions to be approximately additive and sign-reversible. At transfer (Sec.

Method — ALAM: Algebraically Consistent Latent Transitions for Vision-Language-Action Models — Figure 3 : Real-world results on a Piper 6-DoF manipulator. ALAM is compared with $\pi_{0}$ and $\pi_{0.5}$ on four tasks: insert cylinder , insert cube , stack cup , and fold towel .

Experimental Setup

ALAM is pretrained on $11$ action-free video sources, mostly real-robot manipulation data from Open-X-Embodiment O’Neill et al. (2024) and CALVIN Mees et al. (2022) . We use AdamW on $128\!\times\!$H20 GPUs (90G) and run for $39$ epochs (about $4$ days), at which point the pretraining loss has plateaued. The full dataset configuration and hyperparameters are in Appendix A .

The transfer stage is built on $\pi_{0}$ Black et al. (2026) .

Experimental Setup — ALAM: Algebraically Consistent Latent Transitions for Vision-Language-Action Models — Figure 4 : LAM vs. ALAM on the $5\%$ held-out split. (A) Algebraic probes (log $y$, shaded $t{\geq}3k$ not seen during training): (A.1) additivity error and (A.2) reversibility error, with LAM/ALAM ratios at $t{=}5k$ shown in red. (B) Per-horizon $\Delta$ from each model’s own $k$-step score, for direct (top, (a)–(c)) and cumulative (bottom, (d)–(f)) reconstruction. See Sec. 4.1 for definitions and $k$-step scores.

Sources and Demonstrations

Figures are reproduced from Tang et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.