Sriram Gopalan

Building Physical AI @DecisionFacts — teleop & UMI demo data, sim-first VLA training (ACT · SmolVLA · GR00T). Sharing what actually works on real robots.

[Sriram Gopalan @srigopal]

Great to see SO-101 Bench, @h1ckok. Cheap sim you can trust is the unlock for low-cost robots.

A complementary experiment: we replayed real SO-101 teleop (cup nesting, 200 episodes) in open-source MuJoCo. Sim joints stay within ~1° of the real arm on average.

Would nesting make a good next task? Dataset 👇
https://t.co/6XPKLky56l

@huggingface @LeRobotHF

[Trossen Robotics @trossenrobotics]

In @TolokaAI’s tests, GPT-6 Astra controlled our bimanual Trossen AI robot straight from the camera feed. When a grab missed, it tried again. Sometimes it changed its mind halfway through.

Writeup: https://t.co/bur5mJLlhT
Check out our hardware: https://t.co/NBkZjfgez2

[Sriram Gopalan @srigopal]

@vladposi @positronic_ro @lukas_m_ziegler @nebiusai That state timeline is basically subtask segmentation, and the ‘in control → floating’ transition is exactly the failure/recovery pair we try to keep. Do you label it by hand, or derive it from gripper width and object pose?

[Sriram Gopalan @srigopal]

Three things:
1) whether the sim arm matches the real one: if joints track within ~1°, we can trust sim evals for that robot.
2) which episodes are bad: sync drift, dropped frames or joint-limit hits show up as sudden real-vs-sim divergence, often before training would surface them.
3) where sim stops being trustworthy: the biggest gaps are on fast moves and contact, so those are the skills we still validate on hardware.

[Sriram Gopalan @srigopal]

Before we spend a single real-robot hour on a new policy, we replay the teleop demos in MuJoCo.

SO-101 pipe sorting, ep 2: real (left) vs. physics sim driven by the same joints (right). <1° mean tracking error; the biggest gaps are on the fastest moves.

Data 👇 https://t.co/EUz3XgMXE7

[Sriram Gopalan @srigopal]

@RoboSafs01075 Modeling the actual IMU mounting position in training is such an easy thing to skip and such a common sim2real failure. Did you randomize the mounting offset too, or just match it to the real robot?

[Sriram Gopalan @srigopal]

Congrats for your YC26 acceptance. 👏
Adding one to the ‘what breaks when robots leave the lab’ list: reset poses.

Demos recorded from a slightly different start pose teach the policy a task that doesn’t exist in deployment. Curious what’s been the biggest surprise for you so far. We at @DecisionFacts are focussed on that. Let's connect. Thanks!

[Sriram Gopalan @srigopal]

@LeRobotHF @RoboColosseum Real-robot eval at scale is the missing piece. Sim gets you to ‘not obviously broken,’ then the gap shows up in resets and lighting. Will results come with per-episode videos so failures can be traced back to the data?

[Sriram Gopalan @srigopal]

@LeRobotHF This is great. Thanks!
Logging the exact command for every run is the underrated part. When we compare ACT vs SmolVLA checkpoints in sim, half the confusion is which dataset revision a run used. Any plans to log the dataset commit hash next to it?

[Sriram Gopalan @srigopal]

@DisruptConcept Calibration is the half of the tax people see. The hidden half is camera-to-gripper sync. We reject anything over 5 ms of skew, and that check catches more bad episodes than calibration drift does.

[Sriram Gopalan @srigopal]

@axisrobotics No single recipe works for every model’ matches what we see collecting teleop data: operator quality varies within a single session, not just between people.

Did you find per-episode quality labels beat per-operator ones, or is that too noisy to label reliably?

[Robotics Center of Silicon Valley @RoboticsCtr]

LLM vs. WAM: Astra vs. Dream-1. Round 3.

Cup stacking. Same arms. Same task. Zero-shot.

A: 1:06
B: 0:41

Both finish. Which one is the LLM?

A or B? Defend your pick.

#Robotics #EmbodiedAI

[Sriram Gopalan @srigopal]

Meta reportedly built a humanoid prototype with Fourier in about three months.

That timeline says the body is no longer the hard part. The hard part is the thousands of hours of clean demonstrations, with gripper state and contact, that a body needs before it’s useful.

Who solves that first: glasses video, teleop farms, or sim?

[Reuters @Reuters]

Meta builds humanoid robot with Chinese startup Fourier, The Information says https://t.co/3kc9vQ7ve8

[Sriram Gopalan @srigopal]

@YoussefMMSaied Love that you actually ran it. Changing only the seed from the same start is exactly the test I was hoping for. Does it pick each order about equally often, or lean toward one? Comparing that split to the split in the demos would be a great number to report.

[Sriram Gopalan @srigopal]

@KaiboHe Fair enough, and thanks for releasing it. At 4 hours on one 5090, running hundreds of randomized trials per checkpoint looks cheap enough to add on top. If we rerun it, I’ll share what we find.

[Sriram Gopalan @srigopal]

@EitanShaar Agree. A cheap first step is replaying your real teleop episodes in sim. We did that with 200 SO-101 cup-nesting demos and joints tracked within ~1°, which let us probe start states in sim and only re-record where the policy broke. How do you decide which gaps deserve a human?

[Sriram Gopalan @srigopal]

@Frame_Robotics Browser MuJoCo with real HF policies is a great on-ramp for eval.
Feature idea: replay a LeRobot dataset episode next to the policy rollout, so you can tell whether a failure comes from the policy or from the demos (frame drops, gripper jitter).

[Sriram Gopalan @srigopal]

@TheHumanoidHub Three months to a prototype says the body is commoditizing. The slow part is still data: thousands of hours of clean demos with gripper state and contact, which egocentric video alone hasn’t delivered. Who Meta uses for teleop will say more than the hardware.

[Sriram Gopalan @srigopal]

@hankyang94 Physics-aware retargeting is the step most human-video pipelines skip. In our UMI captures the bigger error is contact timing: when the gripper actually closes vs. when the fingers do. Does GRASP adjust timing, or only poses?

[Sriram Gopalan @srigopal]

@BharatJain Nice to see an SO-101 driven zero-shot. Where does most of the cost go, model calls per step? The split that works for us: the big model sets subgoals at around 1 Hz, and a small ACT or SmolVLA policy runs the arm at 30 Hz on the edge.

[Sriram Gopalan @srigopal]

@LeoKharon Latent world models are most useful to us as evaluators, not planners. If one can rank policy checkpoints the same way real rollouts do, that saves a lot of robot hours. Did they report how well its predictions track real success rates?