菜单
Build unified world models using flow-based video generation architecture that handles Text2World, Image2World, and Video2World in a single model, trained on 200M video clips with RL post-training for improved instruction-following and video quality.
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 world-models-video-prediction-robotics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。
正在显示 SKILL.md
| name | world-models-video-prediction-robotics |
| title | World Simulation with Video Foundation Models for Physical AI |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2511.00062 |
| keywords | ["World Models","Video Generation","Robotics","Physical Simulation","Unified Multimodal"] |
| description | Build unified world models using flow-based video generation architecture that handles Text2World, Image2World, and Video2World in a single model, trained on 200M video clips with RL post-training for improved instruction-following and video quality. |
Robotics and embodied AI require understanding how the physical world evolves. Cosmos-Predict is a unified world model that generates videos from text descriptions, images, or previous video clips using a flow-based architecture. The model operates at 1280×720 resolution, handles variable-length sequences, and grounds generation in physical reality through alignment with a physics-aware vision-language model.
The key innovation is unifying multiple input modalities (text, image, video) into a single generative framework, enabling flexible control of world simulation.
Unified Multimodal World Simulation:
1. Design Flow-Based Video Generator
Implement the core generative model using flow matching.
class CosmosFlowVideoGenerator(nn.Module):
def __init__(self, hidden_dim=2048, num_blocks=24):
self.hidden_dim = hidden_dim
# Video encoder/decoder using 3D convolutions
self.video_encoder = nn.Sequential(
nn.Conv3d(, , kernel_size=(, , ), stride=(, , ), padding=(, , )),
nn.ReLU(),
nn.Conv3d(, , kernel_size=(, , ), stride=(, , ), padding=(, , )),
nn.ReLU()
)
.transformer = nn.ModuleList([
nn.TransformerEncoderLayer(d_model=hidden_dim, nhead=, batch_first=)
_ (num_blocks)
])
.video_decoder = nn.Sequential(
nn.ConvTranspose3d(, , kernel_size=(, , ), stride=(, , ), padding=(, , )),
nn.ReLU(),
nn.ConvTranspose3d(, , kernel_size=(, , ), stride=(, , ), padding=(, , ))
)
():
video = video.permute(, , , , )
latent = .video_encoder(video)
latent
():
video = .video_decoder(latent)
video = video.permute(, , , , )
video
():
latent = .encode_video(video)
batch, channels, frames, h, w = latent.shape
latent_flat = latent.permute(, , , , ).reshape(batch * frames * h * w, channels)
block .transformer:
latent_flat = block(latent_flat.unsqueeze()).squeeze()
latent = latent_flat.reshape(batch, frames, h, w, channels).permute(, , , , )
video_recon = .decode_latent(latent)
video_recon
2. Implement Multimodal Conditioning
Support text, image, and video inputs as conditions.
class MultimodalConditioner(nn.Module):
def __init__(self, hidden_dim=2048):
# Text encoder
self.text_encoder = T5EncoderModel.from_pretrained('t5-large')
self.text_projection = nn.Linear(768, hidden_dim)
# Image encoder
self.image_encoder = timm.create_model('vit_large_patch16', pretrained=True)
self.image_projection = nn.Linear(1024, hidden_dim)
# Video encoder (reuse video encoder from generator)
self.video_encoder = VideoEncoder()
self.video_projection = nn.Linear(256, hidden_dim)
def encode_text_condition(self, text):
# text: list of strings
tokens = self.text_encoder.tokenizer(text, padding=True, return_tensors='pt')
embeddings = self.text_encoder(**tokens).last_hidden_state
# Aggregate over sequence: [batch, seq_len, 768] -> [batch, hidden_dim]
condition = embeddings.mean(dim=1)
condition = self.text_projection(condition)
return condition
def encode_image_condition(self, image):
# image: [batch, 3, 720, 1280]
features = self.image_encoder.forward_features(image)
condition = self.image_projection(features)
return condition
def encode_video_condition(self, video):
# video: [batch, frames, 3, height, width]
# Sample key frames to reduce computation
key_frames = video[:, ::4, :, :, :] # Every 4th frame
latents = self.video_encoder(key_frames)
condition = latents.mean(dim=1) # Average over frames
condition = self.video_projection(condition)
return condition
def forward(self, text=None, image=None, video=None):
conditions = []
if text is not None:
conditions.append(self.encode_text_condition(text))
if image is not None:
conditions.append(self.encode_image_condition(image))
if video is not None:
conditions.append(self.encode_video_condition(video))
# Combine conditions
combined_condition = torch.stack(conditions).mean(dim=0)
return combined_condition
3. Integrate Physics-Aware Grounding (Cosmos-Reason)
Use a vision-language model to ground generation in physical understanding.
class PhysicsGroundedGenerator(nn.Module):
def __init__(self, base_generator, vlm_model):
self.generator = base_generator
self.vlm = vlm_model # Cosmos-Reason model
def generate_with_grounding(self, condition, num_frames=16):
# Generate candidate videos
generated_video = self.generator(condition)
# Grade video for physical plausibility using VLM
# Ask: "Does this video show physically realistic motion?"
for frame_idx in range(num_frames):
frame = generated_video[frame_idx]
plausibility_score = self.vlm.classify(
f"Is this frame physically plausible? {frame}"
)
if plausibility_score < 0.5:
# Regenerate this frame with stronger physics constraint
# This is a heuristic; in practice, would use more sophisticated methods
pass
return generated_video
4. Implement RL Post-Training for Quality
Optimize video quality and instruction-following through RL.
def rl_post_training(generator, vlm_model, training_configs, num_steps=10000):
optimizer = torch.optim.Adam(generator.parameters(), lr=1e-5)
for step in range(num_steps):
# Sample conditions (text, image, video)
conditions = training_configs.sample_batch()
# Generate videos
generated_videos = generator(conditions)
# Evaluate video quality
quality_scores = []
for video in generated_videos:
# Metrics: FVD (Fréchet Video Distance), user preference, instruction alignment
fvd_score = compute_fvd(video, training_configs.reference_videos)
instruction_align = vlm_model.classify(
f"Does this video match the instruction: {conditions}?"
)
quality = 0.6 * fvd_score + 0.4 * instruction_align
quality_scores.append(quality)
# Policy gradient update
quality_tensor = torch.tensor(quality_scores)
loss = -quality_tensor.mean() # Maximize quality
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step % 100 == 0:
print(f"Step {step}: Quality {quality_tensor.mean():.3f}")
When to Use:
Hyperparameters:
When NOT to Use:
Pitfalls:
Integration Point: Use as world model for model-based RL—train robot policies in simulation, then transfer to real world.
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
