Zhenyang Chen

Robotics Ph.D at Georgia Tech

[Danfei Xu @danfei_xu]

It turns out most SOTA VLAs are not very steerable: you cannot simply say, “Pick up A. No, sorry, I meant B.”

In @DanielZhenyang 's RSS'26 paper ReSteer, we systematically study this gap, propose a way to quantify steerability, and show how to improve it for any VLA.

[Zhenyang Chen @DanielZhenyang]

Most VLA-steering work happens at inference time. We argue steerability is also a data problem — and the real win is a tool that tells you which data is missing, so you can collect it on purpose. Understand the data, get the behavior.

📄 Paper: https://t.co/s1r4E7SEaj
📷 Project: https://t.co/GR1M6vYnwY

This work cannot be done without wonderful collaborators and advisors. Alan, Liquan, Benjamin, @CelineLinatGT @Yuxiao_Chen_ @siddkaramcheti @danfei_xu

[Zhenyang Chen @DanielZhenyang]

Today's frontier VLAs can do many tasks—but they're far less steerable than advertised. Switch the instruction mid-execution and frontier models often plow ahead with the original task anyway.

🤖In our #RSS2026 paper, we measure steerability, explain why it breaks, and fix it. Introducing **ReSteer** (1/7)

[Zhenyang Chen @DanielZhenyang]

LIBERO sim: +11% steerability across 18k rollouts, beating CAST and other strong baselines — while preserving single-task performance.

Real-world (DROID): 2.2× higher steerability across 4 single-task evals + 3 steering scenarios.

Robots that listen, even mid-execution. (6/7)

[Zhenyang Chen @DanielZhenyang]

State-action and visual gap are the key blockers for leveraging egocentric data. Checkout Egoengine led by @Randle_Liu @ShuoCheng94 on how we tackle this↓

[Yangcen Liu @Randle_Liu]

The web is full of egocentric human videos. But robots can’t directly use them as demonstrations yet.
Meet EgoEngine: From Egocentric Human Videos to High-Fidelity Dexterous Robot Demonstrations, for zero-shot visuomotor learning without real-robot demos.
https://t.co/2DZx9xMPEH

[Zhenyang Chen @DanielZhenyang]

@ShuoYangAIR 向妙动科技发送贺电！

[Danfei Xu @danfei_xu]

A few big updates from https://t.co/87ZgYr4hRb:

1. EgoVerse is coming to RSS'26! @ryan_punamiya will present the work, and @simar_kareer will give a keynote at the Data-Centric Robotics workshop.

2. EgoVerse is expanding with more industry partners, including @microagi, @LightwheelAI, and Trace Labs, with a growing pool of training-validated, permissively licensed data (hitting 10k hours soon).

3. We are building the ecosystem around egocentric intelligence: Through a new partnership with the @nvidia Inception Program, we are connecting data vendors with researchers who can put this data to work.

[Baoyu Li ✈️ RSS2026 @BaoyuLi6]

Egocentric human data is abundant, but human motion is not always positive supervision for robot policy due to embodiment gaps. Naive BC co-training can HURT performance ☹️.

🌟Our key finding in **EgoWAM**: the state-prediction branch of a World Action Model effectively bridges this embodiment gap, enabling robot performance to scale with diverse **in-the-wild** human data.

💡The key question then becomes: what world representation transfers best across embodiments?

👇🏻Let’s take a deep dive into it:
🌐 https://t.co/VnhUs8CFKf

🧵[1/]

[Genesis AI @gs_ai_]

We are back. After one year of quiet building.

Introducing GENE-26.5, our first robotic brain that takes a major step toward human-level capability.

For years, robotics has struggled to learn from the world’s largest and valuable data source: Humans.

Solving it means rethinking the whole stack from the ground up:

- A robotics-native foundation model.
- A 1:1 human-like robotic hand.
- A noninvasive data collection glove for motion, force, and touch.
- A simulator that turns weeks of experiments into minutes.

GENE-26.5 is trained across language, vision, proprioception, tactile, and action. We designed a set of tasks to test how far we can go with this new paradigm.

Fully autonomous, 1x speed, one model, same weights. (Enjoy with sound on)

We are approaching the endgame for robotics.

And this is just a beginning.

[Jason Zhu @GoSailGlobal]

Stanford CS336 上，Tatsu 讲了一节 LLM 架构课，把过去 3 年所有主流 LLM 拆开，看它们的共通模板

结论挺爆：90% 的架构选择已经收敛，你随便挑一个开源大模型，它跟其他模型在这些维度上几乎一模一样

讲师的原话
- 2024 年大家都在 cosplay Llama2
- 2025 年的主题是「怎么训得不崩」
- 2026 年的主题是「怎么扛住长上下文」

下面是 2026 年开源 LLM 的标准模板 你训自己的模型可以直接抄

【架构层 已经收敛的 7 件事】
1）Layer Norm 挪出残差流（pre-norm）
原版 Transformer 把 LN 放在残差里 几乎所有现代模型都挪到外面
原因：keep your residual stream clean 梯度反传更稳

2）RMS Norm 替代 LayerNorm
LayerNorm 的减均值 + 加 bias 那部分实际没怎么帮上忙
丢掉之后 flops 只省 0.17% 但运行时省到 25%
（瓶颈在数据搬运 计算反而次要）

3）所有 bias 项全删
跟 RMS Norm 一个道理 系统层省内存搬运

4）激活函数用 SwiGLU 或 GeGLU
gated linear unit 几乎所有现代模型都用
Llama 系 / Qwen / Mistral 用 SwiGLU
Google 系（Gemma / T5）用 GeGLU
区别极小 选哪个都行

5）位置编码用 RoPE
2024 年之后基本统一了
原理：把每对维度按位置旋转一个角度 让 inner product 只依赖相对位置

6）Transformer block 串联（不是并联）
GPT-J / Palm 试过并联 现在基本被放弃
串联的实现优化得太好了 并联省的那点系统开销不值得损失表达力

7）Layer norm 可以「撒」
哪儿不稳就在哪儿加 LN
attention 之前能加 之后能加 两边都加（double norm）也可以
现代模型很多这样做

【超参数 已经收敛的 5 个数】
1）feedforward 维度 / hidden 维度
- 非 GLU 模型：4 倍
- GLU 模型：8/3 ≈ 2.67 倍（因为 GLU 多一组矩阵 要保持总参数量）
- Llama 系：3.5 倍
- T5 1.0 试过 64 倍 后来 T5 1.1 改回标准 别学

2）head 数 × head 维度 ≈ hidden 维度
几乎所有模型都遵守 T5 是为数不多的例外

3）模型纵横比（hidden / 层数）≈ 100
太深 pipeline parallel 难做
太宽 表达力受限
100 这个数字是系统约束 + 表达力的平衡点

4）vocab size
单语模型：30K 左右（早期 GPT-2 那种）
多语 / 通用模型：100K-200K（GPT-4 / Llama 3 / Gemma 都在这个范围）
现代基本都是后者

5）weight decay
仍然普遍使用
但研究发现它在 LLM 里干的事其实是优化器干预 让你最终能收敛到更深的最优点
跟你想的「防过拟合」没什么关系
所以别因为「单 epoch 不会过拟合」就把它关掉

【稳定性 三个救命 trick】
训练大模型最怕中途 loss 突然飙升 然后 NaN 全军覆没
现代模型用三个 trick 防这件事

1）Z-loss
output softmax 的 normalizer 容易爆
加一个 (log Z)² 的正则项 让 Z 始终接近 1
DCLM / Olmo 都用

2）QK norm
attention 的 Q 和 K 在矩阵乘之前各加一个 LN
让 softmax 的输入永远是单位尺度
multimodal 圈先用起来 现在所有大模型都加

3）Logit soft cap（仅 Google 系）
attention logit 用 tanh 硬封顶
Gemma 2/3/4 都在用 但会损失一点点性能 慎用

【Attention 两个新趋势】
1）GQA（Grouped Query Attention）几乎统一
原版 multi-head 推理时 KV cache 会让算术强度崩到 1/h
GQA 共享 K 和 V 但保留多个 Q
表达力几乎不损失 推理成本砍掉 80%
现在所有要做生产部署的大模型 没有不用 GQA 的

2）局部 + 全局 attention 交替
处理长上下文的新方式
Cohere Command A 起头 现在 Llama 4 / Gemma 4 / Olmo 3 全在用
比如每 4 层有 1 层 full attention 其他 3 层是 sliding window 只看附近的 token
比纯 SSM 更稳 比纯 full attention 便宜得多
（Qwen 3.5 做了变体 把 sliding window 那 3 层换成 SSM）

收尾一句
如果你正在训自己的 LLM，上面这一套就是 2026 年的「默认配置」 不需要重新发明，直接抄

如果你只是想看懂 GitHub 上那些 modeling_xxx.py
这一份足够你不再被术语吓住

[Tairan He @TairanHe99]

GR00T-VisualSim2Real is now open source!

VIRAL and DoorMan are now available with training code, simulation assets, and the full recipe for bringing visual sim-to-real loco-manipulation skills to your own humanoids.

Repo: https://t.co/vgRsCeRG8w

[Zhenyang Chen @DanielZhenyang]

when you ask Codex to optimize doc in codebase, this is what happened:
it first happily deleted the old CLAUDE.md and write this 😂😭
#claude #codex #ai #code #robotics

[Saining Xie @sainingxie]

vision🍌 is here https://t.co/Ued6GGk4Et
if you got into computer vision the way I did, starting with pixel-level labeling tasks like segmentation, edges, depth, or surface normals, you’ll probably feel the same seeing these results -- something big has quietly shifted, and it’s going to change how we approach these problems for good 🧵

[Younghyo Park @younghyo_park]

What's different between these two BC policies? It's the same architecture, training budget, and data collection setup — the only difference is the controller gains!

Controller gains are an understudied design parameter in robot learning. In our new work (w/ @BronarsToni*, @pulkitology), we show how they act as an inductive bias across BC, RL, and Sim2Real transfer, with real consequences on performance. Here's what we found 🧵

* Equal Contribution
📄arxiv: https://t.co/SMYgh7i8cA
🔗website: https://t.co/cLCd1FYCdJ

[Patrick OShaughnessy @patrick_oshag]

My conversation with Sergey Levine (@svlevine).

Sergey is the co-founder of @physical_int -- a company building foundation models that can control any robot to do any task in any environment.

The company's thesis is that generality is more scalable than specialization, meaning that a model trained across many different robots and tasks will ultimately outperform any system built to do one thing well (eg, just wash dishes).

Sergey is a researcher by background, but I think you will appreciate how practical and commercially grounded this conversation is.

We discuss:
- Why changing a diaper will be the last task a robot masters
- The simulation v. real-world data debate
- How multimodal LLMs give robots common sense
- Moravec's Paradox + Robot Olympics
- Why robots can do long-horizon tasks now
- A realistic timeline for robots in our homes

I should note that I am an investor in Physical Intelligence -- I made the investment because I believe it is one of the most important companies tackling the problem of robotics.

Enjoy!

Timestamps:
0:00 Intro
2:39 Defining Physical Intelligence
5:19 The Challenge of Building General Models
6:34 The Stakes and Future of General Purpose Robotics
8:15 Pros and Cons of Humanoid Robots
10:12 Historical Milestones in Robotics Research
15:31 Combining Generative AI and Deep RL
21:24 Moravec's Paradox 25:33 Kitchen Robots
29:30 Simulation vs. Real-World Data
30:48 The Robot Olympics
36:31 The Physiological Reality of Embodiment
38:56 Controversies in the Robotics Community
44:18 What Makes a Great Researcher
48:27 How Businesses Should Prepare for Robotics
54:09 Tracking Progress Through Research Papers
57:02 The Next Step: Mid-Level Reasoning
1:02:00 The Kindest Thing