Aether AI × UC San Diego · Embodied World Models

CD-LAM: Causally Debiased Latent Action Model for Embodied Action Conditioned World Models

A LAM-side causal debiasing method for continuous latent actions — it debiases the latent action space before world-model training, so the action condition carries embodiment motion instead of background shortcuts, with the downstream conditioning format unchanged.

Yufan Wei1,2*, Kun Zhou1†, Lingjun Mao1,2*, Zijun Zhang1, Ziming Xu1, Ziqiao Xi1, Shuang Liang1,2*, Ruobing Han1, Yuchen Yan1, Xinyue Wang1,2*, Fan Feng1, Biwei Huang1
1Aether AI  ·  2University of California, San Diego
*Done during an internship at Aether AI.    Project Lead & Corresponding author: franciskunzhou@gmail.com
0%
action-following error (FDCE)
+0 dB
PSNR over DreamDojo
0×
fewer robot-action training updates
0 h
of debiasing data is enough
CD-LAM pipeline: Stage 1 LAM debiased fine-tuning, Stage 2 ACWM debiased fine-tuning, Stage 3 robot action based training.
The CD-LAM pipeline. Stage 1 (LAM debiased fine-tuning) suppresses action-irrelevant confounders inside the latent action model; Stage 2 (ACWM debiased fine-tuning) trains the world model on the debiased latents; Stage 3 (robot action based training) aligns executable robot actions to the already-debiased space — the conditioning interface stays unchanged.
The Problem

Latent actions are confounded — and world models inherit it

Latent action models (LAMs) let world models learn control from ordinary, action-unlabeled video. But trained only to reconstruct the next frame, a LAM encodes whatever is predictive — embodiment motion, and also background motion, camera shake, and scene appearance. The world model then receives those confounders as part of its action condition: rollouts stay visually plausible while control quietly fails.

The causal picture: one path must go
conditioning Action-irrelevant confounders C background · object identity · lighting Embodiment motion A arm & gripper dynamics z confounded debiased latent action ACWM action-conditioned world model rollout confounded rollout follows action context leakage id_ratio 0.527 · action-norm null ρ 0.711 · held-out track error 20.08 px context leakage id_ratio 0.047 · action-norm null ρ 0.115 · held-out track error 11.50 px
Prediction shortcuts become control failures. A frame pair carries embodiment motion and context. A reconstruction-trained LAM lets context shortcut into the bottleneck (red path), so the world model follows context rather than command. CD-LAM’s training pressures make C→z unprofitable and preserve A→z→Y — toggle above, or watch it cycle.
Reconstruction-trained LAM

Confounded action condition

Latents absorb whatever predicts the next frame:

  • Background & camera motion leak into the latent action
  • Zero action ≠ zero motion — the scene keeps moving
  • Transferred actions fail to steer new scenes
CD-LAM (ours)

Debiased action condition

Latents face embodiment dynamics; context stays in observations:

  • Confounding path cut before world-model training
  • Zero action ⇒ motion largely suppressed
  • Same 32-D latent interface — drop-in for the ACWM
Original episode · context
CD-LAM · do(u=0) FDCE 2.5
DreamDojo · do(u=0) FDCE 28.9
Zero action, nonzero motion — live. Both models receive a zero action; the left tile is what the episode actually did. DreamDojo hallucinates roughly that motion from visual context alone — 56× the residual embodiment motion of CD-LAM (105.5 vs 1.9 px, 14B). More in the rollout gallery ↓
The Approach

Debias the LAM once — everything downstream gets better

Three objectives — embodiment-centric reconstruction, action-centric contrastive learning, and latent space calibration — applied in a lightweight three-stage pipeline, with the downstream conditioning format completely unchanged.

1

Stage 1 · LAM debiased fine-tuning

A short fine-tune on action-unlabeled video focuses latents on embodiment dynamics, pulls action-consistent transitions together across scenes, and anchors identical frames to a zero-transition reference.

1k steps · unlabeled video
2

Stage 2 · ACWM debiased fine-tuning

The action-conditioned world model trains on latent actions extracted by the debiased encoder — inheriting clean, motion-facing controllability at scale.

same architecture & interface
3

Stage 3 · robot action based training

A lightweight action-to-latent mapping aligns executable robot actions to the already-debiased space — so adaptation to real robot control is fast and cheap.

12× fewer updates
Project Video

See CD-LAM in action

A looping highlight reel of side-by-side rollouts — replay, zero-action intervention, and target-action transfer, labels burned in. A narrated overview is in production.

Auto-looping highlight reel · 8 episodes · sound-free. Full narrated overview coming soon.

Results

Cleaner actions in, controllable rollouts out

Evaluated against DreamDojo on EgoDex (latent-action rollouts) and AgiBot (robot-action rollouts and interventions), with identical architecture, latent dimension, and conditioning format — every gain comes from the LAM-side debiasing stage alone.

0% / −0%
FDCE, latent-action rollouts (2B / 14B)
0% / −0%
FDCE after robot-action training (2B / 14B)
+0 / +0 dB
PSNR over DreamDojo (2B / 14B)
0k vs 0k
steps to match the baseline's full budget

FDCE, head to head Foreground Displacement Chamfer Error · px, lower is better · robot-action conditioned

Rollouts · replayed actions
2B
DreamDojo12.63
CD-LAM8.24
−34.8%
14B
DreamDojo11.11
CD-LAM7.73
−30.4%
Residual motion under do(u=0) · should be zero
2B
DreamDojo10.71
CD-LAM5.03
−53.0%
14B
DreamDojo9.36
CD-LAM2.18
−76.7%

Zero-action rows measure spurious embodiment motion when the command says “don’t move” — the direct fingerprint of confounding. At 14B, CD-LAM cuts it to less than a quarter.

ModelFDCE (px) ↓PSNR (dB) ↑
DreamDojo-2B12.6319.85
CD-LAM-2B8.24 −34.8%20.60 +0.75
DreamDojo-14B11.1120.01
CD-LAM-14B7.73 −30.4%21.01 +1.0

Robot-action conditioned rollouts after robot action based training; lower FDCE = better action following.

CD-LAM at a glance: better action following, visual fidelity and efficiency than DreamDojo.
At a glance. Against DreamDojo at 2B and 14B scale, CD-LAM improves action following (a) and visual fidelity (b) while matching the baseline with a fraction of the robot action based training budget (c). The reason (d): reconstruction-trained LAMs leak action-irrelevant confounders into the action condition — CD-LAM cuts that path before world-model training.
Side-by-side rollouts: DreamDojo drifts from the commanded trajectory, CD-LAM tracks it at 2B and 14B.
Same start frame, same commanded actions. DreamDojo's arm drifts off the ground-truth trajectory — and going to 14B doesn't fix it. CD-LAM tracks the commanded motion closely at both scales.
FDCE by action type: CD-LAM beats DreamDojo on all eight action categories; distributions shift toward lower FDCE and higher PSNR.
Better on every action type. (a) FDCE drops across all eight categories — pick, place, push, handover, pour, fold, grasp, release — at both scales. (b) Rollout distributions shift toward lower FDCE and higher PSNR.
Training curves: CD-LAM crosses the DreamDojo full-budget reference within about 2-3k steps and keeps improving.
A debiased space is cheap to adapt. CD-LAM crosses DreamDojo's full-budget reference (dashed) with 6–8% of the robot-action training updates — and keeps improving beyond it. Even 1 hour of debiasing data unlocks most of the gain.
Citation

Cite CD-LAM

@article{wei2026cdlam,
  title   = {Causally Debiased Latent Action Model for Embodied Action Conditioned World Models},
  author  = {Wei, Yufan and Zhou, Kun and Mao, Lingjun and Zhang, Zijun and Xu, Ziming and Xi, Ziqiao and Liang, Shuang and Han, Ruobing and Yan, Yuchen and Wang, Xinyue and Feng, Fan and Huang, Biwei},
  journal = {arXiv preprint arXiv:2607.09185},
  year    = {2026}
}