# Humanoid Transfer and Adaptation of Pretrained WAM

Humanoid Transfer and Adaptation of Pretrained WAM

Full article: https://haiguangboy.com/posts/wholebodywam

## Core Method

Preserve manipulation priors, add humanoid adaptation structure
↳ ① SAF, separate expression, joint generation. Preserve the pretrained manipulation pathway, add an independent whole-body control command stream, and feed it together with future vision into a shared DiT. Manipulation and body commands each have their own semantics while participating in joint modeling.

↳ ② UWBC, unify control semantics first. Use a 56-dimensional interface to organize heterogeneous controller commands: 46 shared dimensions and 10 dedicated residual slots. Different controllers select valid fields according to capability, giving similar body motions a consistent expression convention.

↳ ③ CASA, write coordination requirements into attention. When arm manipulability along the task direction decreases, strengthen the body command's attention to manipulation information. The model obtains both hand manipulation intent and the kinematic signal for when body compensation is needed.

## Key Results

What the adaptation structure brings
· In a six-task simulation comparison using the same additional post-training data, WholeBodyWAM achieved a 91.9% success rate, versus 73.6% for DreamZero-PT. After switching to a capacity-matched single action stream, it dropped to 80.8%, the largest loss among the three structural ablations.

· In real-robot experiments on G1 with eight tasks, 20 trials per method per task: in-distribution / out-of-distribution success rates were 81.3% / 68.8%, versus 57.5% / 40.0% for DreamZero.

· After separately fine-tuning the three controllers SONIC, AMO, and GEAR, the average success rate was 89.2%; cross-controller success-rate variance was reduced by 70% relative to Cosmos-3.

## Related Route Comparison

- Opposite route · [World Models: Should You Imagine at Inference Time?](https://haiguangboy.com/posts/omega-eva) `omega_eva_envision_verify_and_act_with_latent_interactive_world_models_2026_08`: Existing ablations do not isolate the benefit of the world model prediction objective itself contradicts decomposition ablation: removing the imagine-future branch drops 1.4 points and falls below the unrefined baseline, removing the proposal branch drops 2.6 points
- Opposite route · [The First Scaling Law for Zero-Shot Human-to-Robot Transfer](https://haiguangboy.com/posts/dyna-2-scaling-law) `dyna_blog_dyna_2_20260811_2026_08`: Existing ablations do not isolate the benefit of the world model prediction objective itself contradicts route bet: the cause that makes cross-embodiment transfer emerge is world modeling (predicting future video) itself, not just data volume—the WAM architecture outperforms pure VLA
- Same route · [Gemini_Robotics_2_Safety](https://haiguangboy.com/posts/gemini_robotics_2_safety) `gemini_robotics_2_safety_2026_07`: CASA provides kinematic guidance and does not yet constitute an executability guarantee validates empirical finding: there is a clear gap between the model 'understanding constraints' and 'acting according to constraints'
- Same route · [JEPA-WAM: An Innovative Method Makes World Models Better Understand State Changes](https://haiguangboy.com/posts/jepa-wam) `jepa_wam_learning_vision_language_action_policies_with_joint_embedding_world_modeling_2026_08`: The real-robot lead includes post-training differences and cannot all be counted as structural gains validates five ablation groups: joint objective 79.2 versus future-only objective 77.3, preserving spatial structure versus iREPA convolution 74.7
- Same route · [LAWM: Why Action Labels Become a Burden](https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling) `latent_action_pretraining_through_world_modeling_2026_07`: LAWM: Why Action Labels Become a Burden
- Same route · [N0-Foundation: Opening a New Era of Touch](https://haiguangboy.com/posts/mathcaln_0_foundation_towards_the_age_of_tactile_intelligence) `mathcaln_0_foundation_towards_the_age_of_tactile_intelligence_2026_09`: N0-Foundation: Opening a New Era of Touch

## Boundaries

⚠️ Adaptation still has costs
These results are built on a 14B base model and 15,000 whole-body post-training demonstrations. The unified interface still requires field mapping and fine-tuning for the target controller, and directly swapping controllers with the same policy has not yet been verified.

The real-robot lead mixes additional post-training with structural gains and cannot all be attributed to the architecture. CASA is a kinematic soft bias and does not guarantee collision, balance, and contact constraints; the actual deployment latency of 16-step denoising also still needs to be measured.

## Author's Judgment (Not Paper Content, Cross-Paper Synthesis View)

📚 ω0 also focuses on the coordination of locomotion and manipulation. WholeBodyWAM is more focused on humanoid transfer of an existing WAM: connecting pretrained priors, controller semantics, and coordination mechanisms so that existing manipulation capabilities gain a whole-body execution pathway.

❇️ Looking back half a year later, the key is whether the structural gains can be preserved in same-data real-robot comparisons, and how much mapping and fine-tuning cost is still required when changing controllers. These two items directly test the reuse value of this adaptation path.