菜单
Enable robot policies to self-improve through imagination using learned dynamics and value models without physical trial-and-error. Compositional world model separates concerns enabling 35-45% performance gains on contact-rich manipulation.
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 rise-robot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 | rise-robot |
| title | RISE: Self-Improving Robot Policy with Compositional World Model |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2602.11075 |
| keywords | ["Robot Learning","World Models","Reinforcement Learning","Imagination","Policy Optimization","Compositional"] |
| description | Enable robot policies to self-improve through imagination using learned dynamics and value models without physical trial-and-error. Compositional world model separates concerns enabling 35-45% performance gains on contact-rich manipulation. |
Vision-Language-Action (VLA) foundation models excel at semantic understanding but struggle with precise contact-rich manipulation requiring fine motor control. Physical trial-and-error RL is expensive and impractical for real robots. Standard world models struggle to balance action controllability with realistic prediction. The core challenge: learn manipulation policies efficiently without costly physical interaction.
RISE shifts robot learning from the physical world to imagination using a compositional world model with specialized dynamics and value components. Rather than joint training, the approach separates:
This separation enables on-policy RL in imagination, where the policy iteratively improves by proposing actions, simulating consequences, evaluating outcomes, and updating based on advantage signals—all without touching a physical robot.
The compositional world model operates through two specialized components trained with different objectives:
class DynamicsModel(nn.Module):
"""Predicts future observations from current state and action."""
def __init__(self):
# Video diffusion backbone
self.backbone = VideoDiffusion()
def forward(self, observations, actions):
state_action = torch.cat([observations, actions], dim=)
predictions = .backbone.sample(state_action, steps=)
predictions
The value model combines progress and failure signals:
class ValueModel(nn.Module):
"""Evaluates imagined trajectories for advantage computation."""
def __init__(self, vla_backbone):
self.backbone = vla_backbone # Pre-trained VLA
self.progress_head = nn.Linear(hidden_dim, 1)
self.td_head = nn.Linear(hidden_dim, 1)
def forward(self, trajectory):
"""
Evaluate trajectory quality with dual objectives.
trajectory: sequence of observations
Returns: advantage signals for policy optimization
"""
features = self.backbone.encode(trajectory)
# Progress regression: how much closer to goal
progress = self.progress_head(features)
# Temporal-Difference learning: detect failures vs successes
td_values = self.td_head(features)
# Combine signals
return progress + td_values
Self-improving policy optimization loop:
def self_improving_loop(policy, dynamics_model, value_model,
goal_spec, num_iterations=100):
"""
Imagination-based RL loop without physical interaction.
"""
for iteration in range(num_iterations):
# Sample task from curriculum
task = sample_task(goal_spec)
# Generate action proposals with advantage conditioning
optimal_advantages = compute_target_advantages(task)
proposed_actions = policy.sample(
observations=task.initial_obs,
target_advantages=optimal_advantages
)
# Simulate in learned dynamics model
imagined_trajectory = []
obs = task.initial_obs
for t in range(horizon):
action = proposed_actions[t]
obs = dynamics_model(obs, action)
imagined_trajectory.append(obs)
# Evaluate imagined trajectory
values = value_model(imagined_trajectory)
# Compute advantage: difference between proposed and evaluated
advantages = optimal_advantages - values
# Policy optimization step
policy.update(task.initial_obs, proposed_actions, advantages)
yield {
'iteration': iteration,
'mean_advantage': advantages.mean().item(),
'policy_loss': policy.loss.item()
}
When to use: RISE is most effective for:
Key implementation steps:
Task-centric batching strategy:
Expected improvements:
Failure modes to avoid:
Compositional world models enable efficient robot learning by separating dynamics prediction and value estimation. The self-improving loop demonstrates how imagination-based RL can substantially improve policies without expensive physical trial-and-error on real robots.