Pretrain with alignment loss

[l3th4l]

Feature Request: Angular Alignment Loss for Normal Vector Comparison

Context:
We want to implement an angular alignment loss to compare predicted surface normals with their corresponding ideal normals during training. The predicted normals are not directly regressed; instead, they are computed by rotating a predicted action vector by a known angular error.

---

Description

This loss should compute the angular misalignment between:

• n_hat: the error-affected surface normal (predicted),
• n_star: the ideal surface normal (ground truth).

The predicted normal vector is computed as:

n_hat = rotate(A, theta)

Where:

• A is the action (a unit normal vector predicted by the model),
• theta is the angular error vector (e.g., from simulator/environment dynamics),
• rotate(A, theta) applies the rotation defined by theta to A.

---

Loss Function

The final loss should be calculated as:

loss = alpha * alignment_loss(n_hat, n_star)

Where:

• alignment_loss is a function that measures angular discrepancy (e.g., using cosine similarity or arccosine of dot product),
• alpha is a scalar weight to control the contribution of this loss term.

---

Acceptance Criteria

• Implement the rotation function rotate(A, theta)
• Define alignment_loss (preferably cosine-based)
• Integrate the new loss into the training pipeline
• Validate using a toy example where angular deviation can be visualized