[CVPR 2026] JALA: Rethinking Latent Actions for Scalable VLA Pretraining

TL;DR

JALA (Joint-Aligned Latent Actions) is a breakthrough in robotic foundation models that solves the data scarcity problem by learning from 7.5 million human manipulation videos. Unlike previous methods that try to "dream" pixels through reconstruction, JALA aligns internal VLM embeddings directly with latent action dynamics. This allows it to scale across both lab-annotated and unconstrained "in-the-wild" videos, leading to a robot policy that exhibits self-correcting behavior and superior generalization in the real world.

The Problem: The High Cost of "Watching" Humans

To build a generalist robot, we need data. Robot data is rare, but human videos are everywhere. However, we face a Quality-Variety Trade-off:

1. Lab Data: Precise 3D hand tracking, but limited to boring tabletops.
2. Wild Data (Ego4D): Massive diversity, but zero action labels and noisy backgrounds.

Prior attempts like LAPA or World Models try to solve this by reconstructing future video frames. But for fine-grained tasks (like plucking a guitar or using tweezers), pixel reconstruction is too noisy. The model spends all its "brain power" trying to figure out the background wallpaper rather than the subtle movement of a finger.

Methodology: Joint Alignment over Reconstruction

JALA introduces a "Predictive Embedding" approach. Instead of predicting pixels, the VLA predicts latent actions ($z$).

1. The Core Architecture

The framework consists of two main perceivers:

• Latent Action Perceiver (LAP): An Inverse Dynamics Model (IDM) that looks at two frames (start and end) and says, "What action happened here?"
• Latent State Perceiver (LSP): Anchors the VLA's context to the action space.

2. The Innovation: Decoupled EMA Updates

To prevent the model from collapsing or ignoring the action signals, the authors used an Asymmetric EMA update. The LSP (context) and LAP (dynamics) share weights but update differently. This ensures the latent space remains "Action-Portable"—meaning the "knowledge" of how a hand moves is decoupled from "what the room looks like."

Scaling with UniHand-Mix

The researchers curated UniHand-Mix, a staggering 7.5M sample corpus (>2,000 hours of video).

• 5M samples are lab-annotated (precise MANO hand poses).
• 2.5M samples are "in-the-wild" egocentric clips with NO labels.

JALA treats "Wild" videos as alignment targets. Even without labels, the model learns by ensuring its internal state can predict the latent action inferred by the LAP.

Experiments & Results: Real-World Superiority

The results on LIBERO and RoboCasa show JALA consistently beating or matching models with much larger parameter counts.

Key Insight: Robustness to "Visual Shift"

In real-world tests (Franka arm with an Inspire dexterous hand), JALA showed a remarkable ability to ignore environmental changes. When the tablecloth was swapped for an unseen texture in a "Put-Three-Objects" task, JALA's success rate barely dropped (60% $\to$ 58%), while other VLAs like Being-H0 collapsed.

Emergent Self-Correction

Qualitative analysis revealed that JALA-trained robots could "retry" a grasp. If a banana was slipping, the model would retract the wrist and re-adjust the fingers—a behavior never explicitly taught but learned from the diverse dynamics of human "retries" in the wild data.

Critical Analysis & Conclusion

Takeaway: JALA proves that you don't need to reconstruct pixels to learn from video. By focusing on latent action alignment, we can finally tap into the "Dark Matter" of unlabeled internet video for robotics.

Limitations: While JALA handles spatial reasoning beautifully, it still struggles with "affordance reasoning" in complex multi-object scenes (e.g., grabbing a spray bottle by the wrong end). Future work likely needs to integrate better object-centric priors with this latent action space.

Future Outlook: JALA marks a shift from Generative VLAs (reconstructing videos) toward Predictive VLAs (aligning dynamics). This path is computationally cheaper and physically more grounded.