Self Vision (@selfvision) header
Self Vision (@selfvision) profile picture

Entrepreneur

Kepler-452b
Joined July 2020
Photo by Self Vision (@selfvision)Photo by Self Vision (@selfvision)Photo by Self Vision (@selfvision)Photo by Self Vision (@selfvision)Photo by Self Vision (@selfvision)Photo by Self Vision (@selfvision)Photo by Self Vision (@selfvision)Photo by Self Vision (@selfvision)Photo by Self Vision (@selfvision)
"I am the five great elements & also that which is beyond them. I am the entire universe. I am the Veda & that which is beyond the Veda. I am knowledge & ignorance. I am the unborn & the born." अहं पञ्चभूतान्यपञ्चभूतानि अहमखिलं जगत् वेदोऽहमवेदोऽहम् विद्याहमविद्याहम् अजाहमनजाहम्
2
10
79
60,474
The reality gap in robotics will never disappear. The breakthrough isn't a simulator that perfectly replicates reality. It's a robotic learning system that continuously discovers where its understanding of reality is wrong & learns to correct it.
1
5
326
Safety must be intrinsic in the robotics development loop. Improving simulated rewards proves little without independent physical evidence. Uncertainty calibration, regression testing & controlled deployment are essential.
96
The architecture I envision: Physics + learned dynamics → policy training → physical deployment → discrepancy analysis → model correction → targeted scenario generation → retraining → independent physical validation. Every iteration should make the next on more informed.
88
The objective in modern robotics should be to maximize improvement in real-world performance per unit of physical experience and simulation compute. This makes simulation itself an optimization problem: what should we simulate, at what fidelity, & what should we learn from it?
80
The next generation of robotics simulation should be failure-directed. Instead of generating millions of random scenarios, identify the regions of the state-action space where uncertainty & failure risk intersect. Generate training experiences there.
1
95
Uncertainty is more valuable than prediction accuracy alone. When physics-based & learned predictions disagree, the system should recognize the limits of its knowledge. A robot must know when not to trust its world model.
1
139
A world model that predicts a convincing video of a successful grasp may still be fundamentally wrong about the forces involved. Robotics needs action-conditioned world models grounded in physical interactions, geometric constraints & measurable real-world state transitions
78
Foundation world models can provide learned priors about unfamiliar environments, object behavior & possible future states without requiring everything to be manually modeled. But visual plausibility is not same as physical validity
87
Imagine a simulator predicting that a robot's action will move an object 10 cm. In reality, it moves 7 cm. The failure contains information that should improve the simulator itself - but today's robotics dev infrastructure is not designed to handle that effectively.
1
1
228
When a robot fails in reality but succeeds in simulation, we usually treat it as a policy failure. We should also treat it as a failed prediction about the physical world. The difference between predicted and observed behavior is an incredibly valuable learning signal.
115
Domain randomization helps by exposing robots to variations in simulated conditions. But randomizing known parameters cannot compensate for unknown mechanisms. Randomizing friction won't fix an unmodeled contact phenomenon. Robustness to assumptions is not robustness to reality
83
Every simulator is a collection of assumptions about reality: friction, contact, deformation, latency, actuation, perception. A robot trained in simulation doesn't necessarily learn how reality works. It learns how to succeed in a particular mathematical approx of the world.
3
140
The hardest problem in robotics is not the sim-to-real gap. It's closing the sim-to-real-to-sim loop. We have become increasingly good at generating synthetic experiences for robots. We are far less effective at turning physical experiences into better simulation
77
It’s amazing to see so many bozos giving “expert” opinions on Jensen, GPUs and fate of Nvidia. A drunk monkey farting for no reason would be more insightful.
1
3
345
A period of deep uncertainty can act as the most important catalyst to usher in unprecedented clarity. This is true for both individuals and organizations.
1
3
24
900
All the layers of compute for modern robotics need to be reimagined. This includes apps, tooling infrastructure, simulation, training, evaluation & deployment. Today’s VLAs, world & foundation models - all are awfully inadequate for addressing the huge sim-real-sim chasms.
348
Domain randomization chooses a distribution q(θ). Real-to-sim gives that distribution an empirical basis. The goal is to concentrate training on plausible physics while retaining uncertainty the observations have not resolved
4
388
A robot observes yₜ = h(xₜ) + εₜ, not xₜ itself. If calibration treats sensor delay as a dynamics error, it may distort mass or damping to compensate. A simulator can fit the measurements while learning the wrong explanation
4
285
A robot can choose actions that reduce uncertainty about its own physics. The objective becomes u* = argmaxᵤ I(θ; y | u, D). In other words choose the experiment whose observation is expected to reveal the most about the unknown parameters
4
334
Not every parameter error deserves equal attention. If J is task performance, ∂J/∂θᵢ measures local sensitivity to parameter θᵢ. Real-to-sim should spend precision where small errors in physics produce large changes in behavior
5
173