Ziqi Pang

Currently MTS at Amazon FAR, working on robotics. Previously Ph.D. from UIUC, B.S. from Peking University

[Ziqi Pang @ZiqiPang]

Why limit GPT-style visual autoregressive (AR) generation to a fixed spatial order?
🎲 Introducing RandAR — our new framework enabling GPT-style AR models for random-order visual generation!
👉 Website: https://t.co/E5SQ4vemLe
👉 ArXiv: https://t.co/P2qlQfHM06
(1/n)

[Ziqi Pang @ZiqiPang]

Really cool work! We are witnessing more successes of sim-to-real transfer

[Younghyo Park @younghyo_park]

Can a robot assemble tight-clearance parts using only its proprioception? No camera, no tactile sensor, no F/T sensor, no joint torque sensing, no priors over the object geometries or poses: just joint position readings.

It actually can! Blind policies trained in simulation insert pegs, mesh gears, and thread nuts, and the peg policy works zero-shot in the real world. Here's what we found 🧵

Joint work with @BronarsToni, my wonderful co-first author, who's on the job market right now! Advised by @kushalk_ and @pulkitology.

[Ziqi Pang @ZiqiPang]

Good take on this. The answer might differ for the deployment of small/big businesses. For small businesses, I would imagine more motivation to do better co-design. For bigger businesses, e.g., Amazon, this would be harder since a small change in the totes would mean replacing millions or billions of them. In the long run, I am highly confident in generalizing the robot first to fit what we have in the factory.

[Ziqi Pang @ZiqiPang]

A perk of doing robotics at Amazon FAR is to visit a real Amazon shipment center. An absolute experience for embodied AI researcher & people aspiring for robot deployment. Some thoughts:
1. Physical AI requires unique company/organization genes.
2. Go to a factory. It is shockingly clear why corps owning factories physically have deep belief in humanoids, e.g., @Tesla @xiaopenghexpeng @amazon @Xiaomi
3. Physical intelligence is beyond robotics, to the extent that, the shelves, the totes, and the people, are all valuable in delivering values to customers.
4. What kind of intelligence can cover all these layers?
5. This visit is definitely a big inspiration for physical AI researchers.

[Ziqi Pang @ZiqiPang]

I believe dpsk-v4.1-flash-1017 could be the spark of the real wave of robot deployment

[DeepSeek @deepseek_ai]

🎮 Build complete projects from scratch

DeepSeek-V4.1-Flash-1017 uses reinforcement learning with self-feedback to improve end-to-end project development—from game creation to 3D modeling.

See it in action 👇

2/4

[Ziqi Pang @ZiqiPang]

@Daviswh totally

[Ziqi Pang @ZiqiPang]

@embodianalysis Very curious what leads to this leap of performance for Opus 5.5

[Ziqi Pang @ZiqiPang]

@shumochu Nice write and sounds a promising way to build data flywheels 👀

[Ziqi Pang @ZiqiPang]

Absolute DEMO! Every robot company should imagine & create demos like this

[Elon Musk @elonmusk]

Turns out the future does have flying cars

[Ziqi Pang @ZiqiPang]

@jackbot_cs Very insightful 👀

[Ziqi Pang @ZiqiPang]

Check out my lab mate’s work on the recipe of scaling ego data pretraining and transfer to robots

[Jiahua Dong @JiahuaDong]

[7/7] Takeaways

Scaling ego-centric video helps and transfers to robots. But contacts, forces and hidden geometry are never visible in pixels — precisely what simulators represent well. Combining the two may be what’s needed.

Code, checkpoints and our OOD benchmark coming soon.

Project: https://t.co/6Vdz1M8SPm
Paper: https://t.co/GddLekoUyV

[Ziqi Pang @ZiqiPang]

@JiahuaDong Congratulations, Jiahua!

[Ziqi Pang @ZiqiPang]

Just read the blog yesterday https://t.co/cQYC7ynMCK and see this today 👀

[Embodied Analysis @embodianalysis]

Gemini 4 Argon “closely follows” GPT 6 Astra on our Physical Agents Index. As of October 14, Argon ranks second of 18 models at 1,324 points, 25 behind Astra's 1,349.

Here's an example run on the same task: put four bottles into the bin, using handovers when needed. Both runs start from the same layout.

➤ Argon releases all four bottles over the bin within its 69 second recording. The final white bottle is still falling when the footage ends.
➤ Astra's 99.5 second recording ends with the yellow and green bottles still on the table. It releases the pink and white bottles earlier.

Both make repeated gripper adjustments. Argon's white bottle also returns to the table mid-run before a later release. The contrast is in the final object states, not an absence of retries.

One run per model, not a success-rate estimate. Both full recordings are shown at 6× speed; durations above refer to the original footage.

[Ziqi Pang @ZiqiPang]

Gotta ask around my colleagues about dex hands now. Looks very promising!

[nature will @naturewill]

Introducing Natura-1, our first general foundation model for the physical world.

It follows natural-language instructions zero-shot and turns them into action with dexterous hands.

Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience.

On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios.

Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions.

It can also autonomously carry out long and continuous workflows to perform long-horizon executions.

Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands.

Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation.

We expect it to be a long-term research and engineering effort.

[Ning Ding @stingning]

Introducing Natura-1, @naturewill's first general foundation model for the physical world.

It is a single model that follows plain-language instructions zero-shot and carries them out with dexterous hands.

- 50B MoE, pretrained on 1M hours of egocentric human experience.
- Drives 20-DoF dexterous hands from plain language, zero-shot, in a 100ms loop.
- Scaling human data from ~44K to 1M hours cut action MSE on held-out robot data by 67.7%.
- Built on in-house data and compute infra, it handles fine-grained long-horizon tasks and a ~30-minute zero-shot interaction in one continuous take.

The general model exists, and it can only get stronger.

https://t.co/mwTNE6BSK5

[Ziqi Pang @ZiqiPang]

👀

[Embodied Analysis @embodianalysis]

GPT-6 Astra and Claude Opus 5 are nearly tied on Embodied Analysis. Both have 34 wins and 6 losses across 40 votes. Astra holds a 16-point Elo lead, 1222 to 1206.

In this task, every object must be placed in the box with its front side facing left.

At 0:19 in this 2.6×-speed video, Claude Opus 5 has cleared the table and returned both arms to rest. GPT-6 Astra has also cleared the table, but one arm continues moving in and around the box through the final second, with the clip ending at 0:20.

The objects are visible inside the box, but their final orientation is harder to verify. A complete assessment needs to consider placement, orientation, and stopping behavior separately

[Ziqi Pang @ZiqiPang]

Congrats Dejia! Very solid work

[Dejia Xu @Ir1dXD]

Been cooking for a while, and super proud of what our team built 🧑‍🍳👩‍🍳
Check this out 👇 More details dropping soon! 🦾

[Guanya Shi @GuanyaShi]

What is the role of simulation in the era of Astra?

Traditionally, in sim2real transfer, we transfer data or policies, which requires relatively accurate simulation.

What excites me about “agentic sim2real” is that we can also transfer tools, knowledge, and procedures. A roughly correct simulator can give agents a place to test their physical intuition, run experiments, and build reusable capabilities.

That’s the idea behind SimEX, led by @JiahengHu1: develop capabilities through simulated experimentation, then use a few real-world trials to refine both the tools and the simulator.

[Zhenyu Jiang @SteveTod1998]

Excited to share our work, SimEX! Simulation works as a playground for coding agents to develop robot skills, while a few real-world trials help refine both the simulation envs and skills. A step toward coding agents that can act in the physical world!

[Sirui Xu @xu_sirui]

How can humanoids keep getting better on their own? Two loops:

🔁 Training: a data flywheel that grows from its own loco-manipulation skills: https://t.co/7WdD1X73cB

🧩 Test time: robots that behave better as an LLM agent thinks longer and evolves: https://t.co/kasxSkimiq
🧵

[Ziqi Pang @ZiqiPang]

Three months after joining Amazon FAR, and just finished something closely related to deployment use cases (sry for the vague posting) with my excellent teammates. Very excited!

I have learned a lot as a previous multi-modal/agent person entering the field of robotics. My own reflections on bringing a crude model into a reliable & dexterous one:
1. Data flywheel is the bottleneck to scalability, and still requires a lot of research on modeling/infra/hardware.
2. Bringing a model from 90-100 requires deep insights into strong operations
3. Hardware is called “hard”ware for a reason, respect the expertise and process of working together with control & hardware teammates
4. The process of making this use case work make me more confident about making robots work generally within 3-5 years