菜单
Align intermediate visual embeddings in vision-language-action models with 3D geometric representations from pretrained foundation models. Improves spatial understanding and enables faster training (3.8x speedup) without explicit 3D inputs.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill spatial-forcing-vla命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | spatial-forcing-vla |
| title | Spatial Forcing: Implicit Spatial Representation Alignment for VLA Model |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2510.12276 |
| keywords | ["vla","spatial-understanding","representation-alignment","3d-geometry","embodied-ai"] |
| description | Align intermediate visual embeddings in vision-language-action models with 3D geometric representations from pretrained foundation models. Improves spatial understanding and enables faster training (3.8x speedup) without explicit 3D inputs. |
Vision-Language-Action models need spatial understanding for embodied tasks, but explicit 3D inputs (depth maps, point clouds) introduce complexity and sensor noise. Spatial Forcing implicitly teaches spatial comprehension by aligning visual embeddings with 3D representations during training.
Core insight: strong spatial understanding emerges when visual processing layers learn to align with geometric structure. By forcing alignment with lightweight 3D foundation models, VLAs learn spatial reasoning without needing explicit 3D inputs at inference time.
Intermediate Alignment: Rather than modifying final outputs, align intermediate visual embeddings with 3D geometric representations. This implicitly teaches spatial structure at multiple processing stages.
Geometric Guidance: Use pretrained 3D foundation models (3D vision transformers) to provide geometric signals without requiring explicit depth sensors or point cloud inputs.
Efficient Training: The alignment process accelerates convergence, achieving 3.8x speedup over baseline VLA training.
Stage 1: Set Up 3D Foundation Model
Initialize frozen 3D feature extractor:
import torch
import torch.nn as nn
from transformers import AutoModel
class SpatialForcingVLA(nn.Module):
def __init__(self, vla_model_name, geometric_model_name):
super().__init__()
# Load VLA visual encoder
self.visual_encoder = AutoModel.from_pretrained(
vla_model_name,
trust_remote_code=True
)
.geometric_encoder = AutoModel.from_pretrained(
geometric_model_name,
trust_remote_code=
)
param .geometric_encoder.parameters():
param.requires_grad =
.alignment_heads = nn.ModuleDict()
layer_idx [, , ]:
.alignment_heads[] = nn.Linear(
.visual_encoder.config.hidden_size,
.geometric_encoder.config.hidden_size
)
():
batch_size = images.shape[]
visual_outputs = .visual_encoder(
images,
output_hidden_states=
)
depth_maps .training:
torch.no_grad():
geometric_signal = ._depth_to_geometric(depth_maps)
geometric_embeddings = .geometric_encoder(
geometric_signal
)
:
geometric_embeddings =
visual_outputs, geometric_embeddings
():
batch_size, height, width = depth_maps.shape
points_list = []
b (batch_size):
yy, xx = torch.meshgrid(
torch.linspace(-, , height),
torch.linspace(-, , width),
indexing=
)
z = depth_maps[b] / depth_maps[b].()
points = torch.stack([xx, yy, z], dim=-)
points = points.view(-, )
points_list.append(points)
geometric_signal = torch.stack(points_list)
geometric_signal
Stage 2: Implement Alignment Loss
Define spatial forcing loss that aligns embeddings:
class SpatialForcingLoss(nn.Module):
def __init__(self, align_layers=[6, 9, 12]):
super().__init__()
self.align_layers = align_layers
def forward(
self,
visual_embeddings,
geometric_embeddings,
alignment_heads
):
"""
Compute alignment loss across specified layers.
Aligns visual embeddings to geometric structure.
"""
total_loss = 0.0
num_layers = len(self.align_layers)
for layer_idx, visual_hidden in enumerate(visual_embeddings):
if layer_idx not in self.align_layers:
continue
# Project visual embeddings to geometric space
visual_proj = alignment_heads[f'layer_{layer_idx}'](
visual_hidden
)
# Compute cosine similarity alignment
# We want visual to align with geometric structure
visual_normalized = torch.nn.functional.normalize(
visual_proj,
dim=-1
)
geometric_normalized = torch.nn.functional.normalize(
geometric_embeddings,
dim=-1
)
# Alignment: maximize similarity
similarity = torch.mm(
visual_normalized,
geometric_normalized.T
)
# Loss: minimize negative maximum similarity
# (encourages each visual token to align with some geometric token)
alignment_loss = -similarity.max(dim=1).values.mean()
total_loss = total_loss + alignment_loss / num_layers
return total_loss
def compute_full_loss(
model,
images,
actions,
depth_maps,
spatial_forcing_weight=0.5
):
"""
Combine action prediction loss with spatial forcing loss.
"""
visual_outputs, geometric_embeddings = model(
images,
depth_maps=depth_maps
)
# Standard VLA loss: predict actions
action_logits = model.action_head(visual_outputs.last_hidden_state)
action_loss = torch.nn.functional.cross_entropy(
action_logits.view(-1, model.action_vocab_size),
actions.view(-1)
)
# Spatial forcing loss
spatial_loss = spatial_forcing_loss_fn(
visual_outputs.hidden_states,
geometric_embeddings,
model.alignment_heads
)
# Combined loss
total_loss = (
action_loss +
spatial_forcing_weight * spatial_loss
)
return total_loss
Stage 3: Training with Spatial Forcing
Integrate spatial forcing into VLA training loop:
def train_vla_with_spatial_forcing(
model,
train_dataloader,
num_epochs=5,
spatial_weight=0.5
):
"""
Train VLA with spatial forcing alignment.
Depth maps used during training but not required at inference.
"""
optimizer = torch.optim.AdamW(
[p for p in model.parameters() if p.requires_grad],
lr=1e-4
)
spatial_forcing_loss_fn = SpatialForcingLoss()
for epoch in range(num_epochs):
for batch_idx, batch in enumerate(train_dataloader):
images = batch['images'].cuda()
actions = batch['actions'].cuda()
depth_maps = batch['depth_maps'].cuda() # Only during training
# Forward pass
loss = compute_full_loss(
model,
images,
actions,
depth_maps,
spatial_forcing_weight=spatial_weight
)
# Backward pass
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
if batch_idx % 100 == 0:
print(
f"Epoch {epoch}, Step {batch_idx}, "
f"Loss: {loss.item():.4f}"
)
return model
# During inference, depth is not required
def inference_vla(model, images):
"""
VLA inference without depth maps.
Spatial understanding learned through spatial forcing.
"""
with torch.no_grad():
visual_outputs, _ = model(images, depth_maps=None)
action_logits = model.action_head(
visual_outputs.last_hidden_state
)
actions = action_logits.argmax(dim=-1)
return actions
When to Use Spatial Forcing:
When NOT to Use:
Training Strategy:
| Phase | Focus | Depth Required |
|---|---|---|
| Warm-up (Epoch 1-2) | Action prediction only | No |
| Main (Epoch 3-4) | Spatial forcing with weight 0.5 | Yes |
| Fine-tune (Epoch 5) | Reduce spatial weight to 0.1 | No |
Common Pitfalls:
Based on the research at: https://arxiv.org/abs/2510.12276