- Home
- Skills
- Agents & Orchestration
- embodied intelligence expert
Works with the AI Agents you already use
embodied intelligence expert
Expert guidance for embodied AI algorithms, perception-action pipelines, and sim-to-real transfer strategies.
$9
embodied intelligence expert
Example session with this skill installed
I need to train a robotic arm to pick up fragile glassware. I have 100 expert demonstrations but no simulation environment yet. What is the best approach?
- Read your context and instructions
- Compiled the embodied intelligence expert
The task involves vision-based manipulation with expert data, classified as an Imitation Learning (IL) problem with continuous action space and potential for high-precision requirements.
Proposed Strategy: Diffusion Policy for Glassware Manipulation
| Component | Recommendation |
|---|---|
| Core Algorithm | Diffusion Policy (Conditional DDPM) |
| Observation Space | RGB-D + Joint States (Proprioception) |
| Action Space | Action Chunking (Horizon=16) to mitigate compounding error |
| Encoder | Pre-trained ResNet-18 or ViT for visual features |
Implementation Sketch
# Architecture: Conditional Diffusion Policy
# Input: Vision features (v) and Proprioception (p)
# Output: Sequence of predicted actions (a_1...a_T)
class GlasswarePolicy(nn.Module):
def __init__(self):
self.vision_encoder = VisualBackbone(type="resnet18")
self.noise_pred_net = ConditionalUnet1D(
input_dim=action_dim,
global_cond_dim=obs_embedding_dim
)
def forward(self, obs, noisy_action, timestep):
obs_cond = self.vision_encoder(obs)
return self.noise_pred_net(noisy_action, timestep, cond=obs_cond)
Pitfalls to Avoid
Visual Refraction: Glassware causes depth sensor noise. Use heavy data augmentation on depth maps or rely on multi-view RGB if depth is too unreliable.
2.
Overfitting: 100 demos is a small dataset for diffusion. Use EMA (Exponential Moving Average) for weights and strong visual pre-training to improve generalization.
3.
Compliance: Rigid controllers will break glass. Implement a hybrid position/force controller or use an impedance control wrapper for the output actions.
Next steps
- Pre-process the 100 demonstrations into the (Observation, Action) sequence format.
- Train a baseline Diffusion Policy with a horizon of 16 steps and check for MSE convergence.
- Evaluate the policy's ability to handle slight variations in glass position before moving to real hardware.
Connects securely to your tools. The creator never sees your data.
About this skill
Developing algorithms for physical agents often results in a "sim-to-real gap" where policies that work in simulation fail in the real world. This skill provides production-grade guidance for Embodied Intelligence (具身智能), helping developers design and optimize algorithms for perception, planning, and control in robotics.
What it does
- Algorithm selection recommends specific architectures like PPO, SAC, Diffusion Policy, or ACT based on your task constraints.
- Problem framing classifies tasks by observability, action space, and learning signals to ensure correct modeling.
- Sim2Real strategy provides checklists for domain randomization, system identification, and robust policy design.
- Debugging & optimization identifies root causes for common failures like reward hacking, compounding errors, or policy drift.
- Bilingual support delivers technical guidance in both English and Chinese with consistent terminology mapping.
How it works
- Classify the problem using the Core Algorithm Decision Framework to define the state space and learning requirements.
- Select the optimal learning or planning method based on available data, such as expert demonstrations or sparse rewards.
- Draft the implementation architecture using provided pseudo-code or architecture sketches.
- Audit the proposed pipeline against the Sim2Real checklist to prepare for physical deployment.
Frameworks & tools
This skill covers modern embodied AI architectures including Vision-Language-Action (VLA) models, Diffusion Policy, Action Chunking Transformers (ACT), and Open X-Embodiment datasets. It references industry standards like RT-1/2, Octo, OpenVLA, and robust RL libraries.
Why this beats prompting it yourself
Generic AI often suggests outdated or overly academic robotics papers that fail in real-world contact-rich tasks. This skill uses a specialized decision matrix to prioritize sample-efficient, production-hardened methods like Diffusion Policy for manipulation or IQL for offline learning, specifically addressing the compounding error and distribution shift problems inherent in physical AI.
Use cases
- Designing a vision-based manipulation policy for transparent or complex objects.
- Developing a robust walking policy for legged robots using domain randomization.
- Implementing offline reinforcement learning from existing logged robot datasets.
- Fine-tuning Vision-Language-Action models for specific downstream robotic tasks.
Known limitations
Requires specific problem parameters (observability, action space) to give accurate recommendations. Does not provide hardware-specific drivers or circuit design.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Requires specific task parameters to provide accurate algorithm recommendations. Does not provide hardware-level drivers or electrical engineering schematics.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 22 days ago
Permissions
No special permissions declared or detected