tags:
- robotics
- humanoid
- reinforcement-learning
- double-dqn
- pytorch
HumanoidTTT — Consolidation Policy
Initial Double-DQN consolidation weights for HumaniodTTT: Test-Time Capability Reuse for Efficient Humanoid Control.
Model
The policy scores retention actions for a finite-capacity capability store. Each
action has a 71-dimensional consolidation feature and is scored by a shared
71 → 128 → 64 → 1 ReLU network with 17,537 parameters. When the ten-slot store
is full, the actions are SKIP or REPLACE(j). Online Double-DQN updates use the
fraction of subsequent requests with successful reuse between consecutive
full-store qualified-miss decisions.
Release files
| File | Contents |
|---|---|
consolidation_policy.pt |
Initial FP32 PyTorch scorer state dictionary, seed 83001. |
config.json |
Architecture, feature layout, action mapping, online settings, and weight checksum. |
SHA256SUMS |
File checksums. |
This checkpoint is the starting point before online adaptation. It does not contain a prefilled capability store or optimizer/replay state.
Usage
Install the code and download dependencies:
git clone https://github.com/AIGeeksGroup/HumaniodTTT.git
cd HumaniodTTT
pip install -e .
pip install huggingface_hub
import torch
from huggingface_hub import hf_hub_download
from humanoid_ttt import ConsolidationPolicy
path = hf_hub_download("AIGeeksGroup/HumanoidTTT", "consolidation_policy.pt")
weights = torch.load(path, map_location="cpu", weights_only=True)
policy = ConsolidationPolicy(weights, capacity=10, online_learning_rate=3e-5, gamma=0.95)
See the GitHub README for entry applicability, execution feedback, and update interfaces. Pin the model revision and code commit when reproducing experiments.
Dependencies and scope
The 45-dimensional A2 entry-applicability module uses state features and geometric certificates; it has no separate neural-network checkpoint. Applications supply motion-specific certificates, qualification evidence, and the execution runtime. Download frozen generation and tracking models from OMG and HoloMotion.
The released scorer weights alone do not constitute an end-to-end robot controller. Successful loading verifies network compatibility, not complete experimental reproduction.