Pengcheng Wang (@Outsider_pc) header
Pengcheng Wang (@Outsider_pc) profile picture

Ph.D. student @berkeley_ai | Prev. Intern @meta @amazon Transferable and Scalable RL for Robotics. Do mathematically beautiful research trivial in hindsight.

Berkeley, California
Joined February 2025
Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)Photo by Pengcheng Wang (@Outsider_pc)
Real-time Chunking (RTC) is designed to enable smooth asynchronous execution of flow-matching policies. However, it has some critical limitations: its inpainting-based async execution capability comes from inference-time corrections rather than the base policy, yielding little pre-training benefit, specific fine-tuning for better performance (e.g. training-time RTC), heuristic guidance, and extra computation that inflates the latency. In this work, we observe that discrete diffusion policies, which generate actions by iteratively unmasking, are natural asynchronous executors that resolve all limitations at once, being simpler to implement, faster at inference, and better at execution. Paper: arxiv.org/pdf/2604.25050 Code: github.com/outsider86/Discre… Website: outsider86.github.io/Discret…
8
17
6
123
89,225
Pengcheng Wang retweeted
We’re happy to introduce StarVLA, an open-source codebase to help you develop and experiment with your own VLA models. Nearly 4,000 GitHub stars, with 60+ community projects building on it. Models, recipes and benchmarks: starvla.github.io/
8
6
7
44
193,205
Pengcheng Wang retweeted
SuperMimic has won the CORL’s best paper award. Thanks for everyone’s support!
SuperMimic is open-sourced during my talk at CORL. Have fun with it: github.com/HybridRobotics/su…
79
64
23
1,233
75,820
Pengcheng Wang retweeted
Introducing ⚡️Flash ALOHA: Learning Fast and Precise Manipulation! With just 50 demonstrations, Flash ALOHA learns to autonomously complete dynamic and precise tasks at near human speed. Everything open sourced! (1/11) Team: @Qi_Wu577 @YihuaiGao @chelseabfinn @zipengfu – at Stanford, CA
46
119
34
924
119,112
Pengcheng Wang retweeted
Gemini Robotics 2.5 is here. 🤖 Our latest physical AI is able to learn hundreds of everyday skills - from folding laundry and serving pastries to autonomously sorting car parts on the factory floor. Here’s a snapshot of its capabilities 🧵
378
573
407
4,661
498,441
We were presenting DiscreteRTC this afternoon at 3rd floor, #007. Welcome to chat!
Real-time Chunking (RTC) is designed to enable smooth asynchronous execution of flow-matching policies. However, it has some critical limitations: its inpainting-based async execution capability comes from inference-time corrections rather than the base policy, yielding little pre-training benefit, specific fine-tuning for better performance (e.g. training-time RTC), heuristic guidance, and extra computation that inflates the latency. In this work, we observe that discrete diffusion policies, which generate actions by iteratively unmasking, are natural asynchronous executors that resolve all limitations at once, being simpler to implement, faster at inference, and better at execution. Paper: arxiv.org/pdf/2604.25050 Code: github.com/outsider86/Discre… Website: outsider86.github.io/Discret…
1
4
171
Pengcheng Wang retweeted
After OpenAI's 722 math manuscripts: Why was math the first field AI blew up? When proofs are no longer scarce, what is? Who might end up happier? 24 Q&As from a Peking University seminar, in Chinese. English version on Substack (link inside).
2
9
49
7,674
Pengcheng Wang retweeted
Introducing FlashState ⚡: an inference system that flexibly manages the prefix cache of hybrid LLMs, designed for agentic workloads. On GLM 5.3 Flash, FlashState more than doubles vLLM's throughput and provider profit, while cutting latency by 3.7× and user cost by 2.5×.
80
116
32
1,008
169,152
Pengcheng Wang retweeted
Wanted to get into manipulation. Decided to first build a full stack that scales. Plan: copy @GeneralistAI's gripper. Fun weekend project. It was not. So @t_k_233 and I documented every design choice and lesson along the way. Meet Specialist: fully open-source hardware, data pipeline, training, and deployment code. specialist-ai.github.io/
47
113
37
906
103,679
Pengcheng Wang retweeted
Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO
426
762
526
6,564
2,592,386
Pengcheng Wang retweeted
Both policies: 25/25 successful egg transfers. Different contact quality: short drops vs. gradual weight transfer. I’ll present our robot–object–environment perspective at #CoRL2026’s Beyond Binary Success workshop, Nov 12 in Austin. Come chat about tactile manipulation!
1
3
13
380
?
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
66
Pengcheng Wang retweeted
🦾 We're generating dexterous robot demonstrations from human data at scale 🌟 Check out our work at Meta Reality Labs Research 🔥 Generative Neural Retargeting for Human-to-Robot Dexterous Manipulation 🌐 Website: yy-gx.github.io/GNR/ 📚 Report: arxiv.org/abs/2610.12440
7
27
6
270
18,972
Pengcheng Wang retweeted
The bitter lesson in 26 words: Don’t be distracted by human knowledge, as AI has been historically. Instead focus on methods for creating knowledge that scale with computation, like search and learning.
138
999
152
7,554
628,388
Pengcheng Wang retweeted
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.
261
826
288
7,380
1,048,705
Pengcheng Wang retweeted
Three years ago, training a walking policy for this 700-muscle humanoid took 2 days, using a CPU simulator and SAC. That was for a single motion. Today, with SuperMimic and MJWarp, it learns to track 1.5 hours of motion capture in 4 hours on one 5090. muscle-supermimic.beanpow.to…
14
29
7
280
20,913
Pengcheng Wang retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
134
796
156
6,579
888,531
Pengcheng Wang retweeted
We gave tactile sensors goosebumps and helped robots feel friction. 🤖🦾 👉 Using ordinary cabinet bumpers. Meet TacGooseBumps (TacGB), a passive add-on that sticks directly onto existing normal-only tactile sensors, with no changes to their electronics. How does it work? Under tangential loads, the bumps tilt and redistribute pressure, creating patterns that a robot policy can learn from directly, with no explicit force reconstruction. Across four manipulation tasks, TacGB improves insertion success by up to 36 percentage points, enables gentler egg placement, and produces more continuous, straighter drawing. 🌐 Project: jeffwli.github.io/tacgb/ 📄 Paper: arxiv.org/abs/2609.34006
2
5
1
22
10,562
Pengcheng Wang retweeted
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×. Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours. Project: supermimic.github.io Demo: supermimic.github.io/demo/
98
319
124
1,925
292,291
Pengcheng Wang retweeted
Common belief: Robot policies need continuous action heads; discrete actions are 𝐥𝐞𝐬𝐬 𝐧𝐚𝐭𝐢𝐯𝐞 𝐚𝐧𝐝 𝐥𝐞𝐬𝐬 𝐩𝐫𝐞𝐜𝐢𝐬𝐞. ❗We find policies w/ the simplest discrete representations 𝐦𝐚𝐭𝐜𝐡 𝐨𝐫 𝐛𝐞𝐚𝐭 continuous ones—in sim & real, in single-step & diffusion/flow settings. Here's why 🧵
13
49
12
289
31,293
Insightful look into the action design (, and another reason you should try discrete diffusion/flow since we have native DiscreteRTC🥰)
Common belief: Robot policies need continuous action heads; discrete actions are 𝐥𝐞𝐬𝐬 𝐧𝐚𝐭𝐢𝐯𝐞 𝐚𝐧𝐝 𝐥𝐞𝐬𝐬 𝐩𝐫𝐞𝐜𝐢𝐬𝐞. ❗We find policies w/ the simplest discrete representations 𝐦𝐚𝐭𝐜𝐡 𝐨𝐫 𝐛𝐞𝐚𝐭 continuous ones—in sim & real, in single-step & diffusion/flow settings. Here's why 🧵
1
106