Writing posts on this blog — math, code, callouts, images

A short reference for myself (and anyone else writing here) on the syntax this Chirpy-themed blog supports. Add this front-matter to your post:

---
title: "Your title"
date: 2026-04-26 12:00:00 +0800
categories: [Top, Sub]
tags: [tag1, tag2]
math: true        # only if you use LaTeX
mermaid: true     # only if you use mermaid diagrams
pin: true         # pin to the top of the homepage
image:
  path: /assets/img/cover.png
  alt: cover
---

Math (LaTeX)

Inline: $\mathcal{L}(\theta) = \mathbb{E}{\tau \sim \pi\theta} [R(\tau)]$.

Display:

\[\nabla_\theta J(\theta) \;=\; \mathbb{E}_{\tau \sim \pi_\theta}\!\left[\sum_{t=0}^{T} \nabla_\theta \log \pi_\theta(a_t \mid s_t) \, A^{\pi}(s_t, a_t)\right]\]

The classic policy-gradient identity — the workhorse behind REINFORCE, A2C, PPO, and most of what we run in robot RL.

Code blocks with line numbers

import torch
import torch.nn as nn

class TinyPolicy(nn.Module):
    def __init__(self, obs_dim: int, act_dim: int, hidden: int = 256):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(obs_dim, hidden), nn.SiLU(),
            nn.Linear(hidden, hidden),  nn.SiLU(),
            nn.Linear(hidden, act_dim),
        )

    def forward(self, obs: torch.Tensor) -> torch.Tensor:
        return torch.tanh(self.net(obs))

Shell snippets render fine too:

# launch a training run
python train.py --config configs/ppo_humanoid.yaml --seed 42

Callouts

Useful info — links, references, pointers to docs.

A tip from experience — don’t forget to seed your dataloaders.

Warning — this scales O(N²) in batch size; cap N or you’ll OOM.

Don’t do this on a production cluster without a dry run.

Mermaid diagrams

flowchart LR
    A[Raw demos] --> B[Tokenize]
    B --> C[Pretrain VLA]
    C --> D{Eval pass?}
    D -- yes --> E[RL fine-tune]
    D -- no  --> B
    E --> F[Deploy on robot]

Images

![alt text](/assets/img/example.png){: width="600" }
_Caption goes here._

Tables

Method	Real-time?	Sample efficiency	Wall-clock
BC	✅	high	low
PPO	❌	low	high
Diffusion Policy	✅	medium	medium

Footnotes & definitions

The replay buffer¹ stores transitions for off-policy learning.

---

That’s the kit. Now go write something.

1. A FIFO data structure of (s, a, r, s', done) tuples, typically sized 1M for continuous control. ↩︎