David Fang (@dlfangda1) profile picture

Co-Founder and CTO @shepherd_si | Prev: @DynaRobotics, @Cruise, @UMich @sjtu1896

Joined March 2022
Photo by David Fang (@dlfangda1)Photo by David Fang (@dlfangda1)Photo by David Fang (@dlfangda1)Photo by David Fang (@dlfangda1)Photo by David Fang (@dlfangda1)Photo by David Fang (@dlfangda1)Photo by David Fang (@dlfangda1)
David Fang retweeted
Today we're launching Petrichor (YC F26), a wet lab for frontier AI labs. We run experiments in parallel and return the data from every step.
73
33
28
374
84,631
David Fang retweeted
What if your agents could improve themselves and use fewer tokens as they run? The labs aren't incentivized to do this, so we built ourselves, and now we're open-sourcing our entire self-improving agent platform and control plane. And it works with Claude/Codex/Pi agents.
8
15
4
57
141,340
At Shepherd Robotics, we focus on deployment: getting robots to do useful work in customer workflows. Today, we’re sharing the technical work behind Shepherd-0, our physical skill learning system. shepherdrobotics.ai/blog/she…
31
6
2
95
107,748
With few-shot adaptation, Shepherd-0 reaches state-of-the-art performance among the open-weight models evaluated in our setup. Multi-head supervision powered by Shepherd Adapt helps on more challenging, high-mix tasks. All demonstrations come from initial data collection, without DAgger data. High-mix tasks remain difficult, and gains vary by task.
1
93
These learned physical skills become part of the overall robot system. shepherd-0-reasoning maintains memory and task status, coordinates SLAM-based navigation and manipulation, and asks people for help or confirmation when needed. Production reliability still requires human iteration. Next, we’re building toward continuous learning so robots can improve from execution without manual intervention.
77
David Fang retweeted
Fine-tuning VLAs just got so much easier. Today, we're releasing Training on Dreamscale: 1) Pick a base model (e.g. π0.5) 2) Import your dataset (from HF or local) 3) Start training
25
16
8
158
12,849
David Fang retweeted
Using Claude Code or Codex agents to automate a business workflow is like asking Newton or Einstein to solve it from first principles every time. The AgentRun harness we built is more like an experienced analyst: it learns on the job, takes notes, and modifies its workflow after every run. That means runs get faster and cheaper over time.
Day 4 of open sourcing parts of our harness. Today we made significant improvements to our workflow DSL. We are using it in our self-improving loop - turning agentic loops into more deterministic recipes. With these improvements (a lot of @typesafe's Jev in routing, classification, and workflow composition), we have been able to reduce the cost of agents significantly without sacrificing performance. I meant to open source the learning loop as well, but removing the business logic from @grepdotai still needs work. Enjoy the demo tho : )
18
10
1
44
7,983
Excited to share what we’re building: a physical agent system powered by our own foundation models, tackling highly skilled work that’s physically demanding even for experienced technicians. Proud of this team. More soon!
Introducing Shepherd Robotics (YC F26). We build robots for skilled physical work in data centers and advanced manufacturing—from hardware testing and inspection to wiring and assembly—with qualified technicians supervising.
4
1
13
869
Recently talked to someone who's worked with several well-funded robotics companies. After numerous pilots, they've decided to pause. Not because the tech wasn't impressive, but because they couldn't see a clear path to ROI. Meanwhile, robotics funding has never been hotter. Interesting disconnect between what excites investors and what customers are willing to pay for.
2
165
Would be really exciting to mid-train a robotics model from Astra. The strongest foundation models are starting to look increasingly robotics-native out of the box.
GPT-6 Astra scored 95% on a robot control task, up from Fable 5.1's 40%, with 6.2x fewer output tokens at 2.3x lower cost. 🧵
2
16
1,990
Traditional robotics is good at executing predefined tasks. Physical AI today is getting good at executing structured workflows with flexibility and variation within the steps. But a lot of the most valuable physical work looks different: new situations, new workflows, new structures, with little or no prior demonstration. We’re still pretty far from that.
4
307
Love to see the team keep pushing the frontier. Learned something new about unblocking scaling laws with human data from the deep analysis here. Thanks for sharing so openly!
Today we are introducing Dyna-2, a world-action model pre-trained on one million hours of human video. At this scale, for the first time, we discovered several new scaling laws: • world-action models exhibit scaling law on human data across four orders of magnitude, from 1000 to 1,000,000 hours, • this human data scaling law implied a scaling law on never seen robot data, • both data and objective matter; world modeling and scaling on video data are essential for cross-embodiment scaling transfer to emerge 🧵
10
518
One thing I’m learning while exploring robotics: There are endless problems robots could solve. The hard part is finding the ones painful enough, repetitive enough, and valuable enough that someone actually wants the robot today.
1
5
455
The World Cup 2026 schedule is honestly confusing when it comes to figuring out which teams would meet each other in the knockout stage, and I ended up buying the wrong ticket by mistake lol So we build a prediction site with agent:
pick group results, simulate the knockout bracket, see which teams face each other along the way, crown a champion, and share your prediction. Click “info” on any simulated matchup to see mainstream ticket links for that game. Check it out here: kunda.dev/worldcup2026/ #WorldCup2026
1
1
236
It’s wild how fast robotics software development is becoming with coding agents. Started a hobby robot project from scratch, and in one night I had the data collection / training / inference / visualization / runtime stack running and got the policy working. But robotics is still hard for the same reason it has always been hard: the physical world. Most of the time went into assembling the SO-ARM101 and other components, waiting for 3D-printed parts, collecting real robot data, and dealing with hardware details.
1
4
410
Again robotics software sucks. Its so bad. Part of what makes it so bad -- and what makes stuft like ROS useful -- is the amount of boring middleware code you need to write between sensors, processes, etc. Coding models make it so much easier to work with, its wonderful.
Replying to @Ken_Goldberg
I asked Codex to set up ROS middleware, configure a CSI camera, benchmark Gemma 4 models on my Jetson Orin Nano, adapt an OpenClaw-style runtime for VLM + reasoning (“Robotclaw” as I call it), and even build an iOS app to stream LiDAR, camera, GPS, and IMU data from an old iPhone to my robot rover. It’s honestly wild how capable these coding agents are now, and how much time they save. I even “write” way more tests now because the marginal cost is so low.
28
38
4
468
52,637
David Fang retweeted
𝗦𝗵𝗶𝗿𝘁 𝗳𝗼𝗹𝗱𝗶𝗻𝗴: 𝟳𝟱.𝟯 𝘀𝗲𝗰𝗼𝗻𝗱𝘀 𝗶𝗻 𝗱𝗲𝗺𝗼𝗻𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻, 𝟭𝟴.𝟵 𝗶𝗻 𝗲𝘅𝗲𝗰𝘂𝘁𝗶𝗼𝗻. 𝗦𝗮𝗺𝗲 𝗽𝗼𝗹𝗶𝗰𝘆. 𝗡𝗼 𝗿𝗲𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴. The instinct, when a robot runs slowly, is to reach for a better model. Usually that isn't where the problem is. Assuming you already have a trained policy (which is its own unsolved problem, and not one we should pretend is behind us), teleoperation data is captured at human speed because hardware limits and delayed control feedback force it to be. The policy learns the behaviour. It also learns the tempo. Naive acceleration breaks the system long before it breaks the model. The real bottleneck sits in the latency stack. Cameras capture frames 55ms before the timestamp. Joint position readings arrive 50ms late. Tracking controllers lag by 150ms. These delays compound during execution, producing a gap between what the model expects and what the hardware actually does. Push the speed without compensating, and the trajectory goes jerky, the arm oscillates, and success rates collapse. @Dexmal_AI's Realtime-VLA V2 paper treats this as a system problem rather than a model problem. Three layers sit around the unchanged policy: a calibration pass that measures and aligns the real delays against training data, trajectory post-processing that uses quadratic programming to smooth action chunks and pre-amplify commands against control loop lag, and a human-in-the-loop speed adaptation head that learns when to accelerate on low-risk segments and slow down for precision-critical ones. The paper reports execution close to human tempo across physical tasks. PCB placement: 89.5s to 37.8s. Pick-and-latch: 98.6s to 42.6s. Shirt folding within a fraction of a second of the human operator. Real hardware, not simulation. What I find useful about this isn't the raw numbers. It's where the optimisation happens. The policy is untouched. The win comes from infrastructure most teams treat as an afterthought: time calibration, trajectory smoothing, adaptive execution. There's a broader point worth drawing out. The field spends a lot of attention on model architecture and much less on the synchronisation, calibration, and execution layers underneath. That imbalance is how teams end up debugging six months of system instability that has nothing to do with the model they trained. Data is still the harder problem, and I don't want to pretend otherwise. But once you have a policy worth running, whether or not it runs well is a systems question. 𝗠𝗼𝘀𝘁 𝗼𝗳 𝘄𝗵𝗮𝘁 𝗹𝗶𝗺𝗶𝘁𝘀 𝗿𝗼𝗯𝗼𝘁 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 𝘁𝗼𝗱𝗮𝘆 𝗶𝘀𝗻'𝘁 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹. 𝗜𝘁'𝘀 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴 𝗮𝗿𝗼𝘂𝗻𝗱 𝗶𝘁. Paper and project page in comments ⬇️ #RobotLearning
4
12
118
10,725