Aarav Kothari

18 | cs (comp eng) + math @ stanford
juga bonito

[Aarav Kothari @aarav_kothari_x]

Day 3/72 in robot intelligence. Just attended a guest lecture by the goat Howie Choset, one of the pioneers of modern robotics.

For nearly three decades at CMU, he's helped push snake robotics from fundamental research into operating rooms, disaster zones, and other environments conventional robots struggle to reach.

A few things that stuck with me:
- Classical robotics is still incredibly relevant. Much of the robotics deployed in high-stakes environments relies on geometric modeling and motion planning. There's invariable potential in robot learning, but also a lot we can accomplish with the tools we already have.
- Snake robots are a fascinating example. With dozens of coupled degrees of freedom, controlling individual joints is relatively straightforward. Coordinating them to move together in non-uniform environments is much harder. Choset's work tackles this through gait design, motion planning, and exploiting contact with the environment for propulsion.
- Designing hardware so that the remaining challenges become software problems. Mechanical complexity is expensive to iterate on while algorithms can be refined continuously.

He also shared his lab's recent deployment alongside the Venezuelan Red Cross following two major earthquakes. A reminder of just how far robotics has come, and how far it still has to go.

[NVIDIA Robotics @NVIDIARobotics]

We’re excited to announce the official release of NVIDIA Isaac Lab 3.0. 🎉

Isaac Lab 3.0 brings a new modular and lightweight architecture along with major new capabilities across the robot learning workflow: multi-physics and multi-rendering support, training at scale, and streamlined deployment on real hardware.

With Newton integration, the new OVRTX renderer, RL post-training, and LEAPP deployment, there’s a lot to explore.

Here’s a look at what’s new 🧵

(1/7)

[Aarav Kothari @aarav_kothari_x]

Incredible resource for anyone getting into robotics or physical AI. Really cool to see classical robotics and modern learning based approaches taught together.

The exercises have the highest value, will definitely take advantage of https://t.co/REShcJ1kfc

[Marco Pavone @drmapavone]

Ever wondered what it’s like to learn robot autonomy @Stanford? Now you can.

Today, we’re opening up the robot autonomy teaching materials that my collaborators and I have been developing over the past several years—for anyone, anywhere, completely free.

At the heart of this effort is Principles of Robot Autonomy (PoRA), a textbook inspired by the Stanford robot autonomy course series I’ve been teaching since 2017.

Robot autonomy has evolved from a research discipline into a technology deployed in the real world. Building capable autonomous systems requires more than understanding individual algorithms: it requires understanding how perception, state estimation, planning, control, and learning work together in a complete autonomy stack.

That’s what we set out to teach with PoRA: bringing classical robotics and modern AI together in one coherent framework, grounded in both theory and practice.

Coming soon in print from Cambridge University Press, Principles of Robot Autonomy is also freely available online and comes with a comprehensive collection of open educational resources:

📖 Textbook, https://t.co/wRvXr4VCIP — Foundations of robot autonomy, from classical methods to modern AI-based approaches.

🎥 Video lectures, https://t.co/Yw3P1aGzBG Lectures from my Stanford courses on robot autonomy and on optimal and learning-based control.

💻 Interactive exercises, https://t.co/kyPV8AnyB3 — Hands-on Python implementations and Jupyter notebooks.

📝 Teaching materials, https://t.co/a6MqqYAcEN — Lecture slides and resources for instructors.

🌐 Explore all PoRA resources, https://t.co/oAsjrMJxZT — Including Ask the Book, an AI study assistant integrated throughout the textbook to answer questions, explain derivations, and connect ideas across chapters.

📣 Read the Stanford announcement: https://t.co/f6BBBZZQmg

Our goal is simple: to make high-quality education in robot autonomy accessible to students, researchers, engineers, educators, and anyone curious about how autonomous robotic systems work.

I believe that sharing knowledge is fundamental to the mission of a university. Autonomous robotics—and Physical AI more broadly—are among the defining technological frontiers of our time. We want to help more people understand these technologies, build with them, and advance the field.

A huge thank you to my co-authors, @DanieleGammelli, Joseph Lorenzetti, @katielulula, and @GioeleZardini, and to everyone who helped bring this project to life.

Explore the resources, use them in your own learning or teaching, and please share them with anyone who might find them useful!

#Robotics #PhysicalAI #ArtificialIntelligence #AutonomousSystems #Education

@Stanford @StanfordEng @StanfordAILab @StanfordHAI

[Aarav Kothari @aarav_kothari_x]

@praneethsamine Great point. Interesting how the integral term can eliminate steady state error from constant loads (gravity), though too much integral gain can introduce overshoot/windup. Appreciate you adding this!

[Aarav Kothari @aarav_kothari_x]

Day 2/72 in robot intelligence.

My experience with robotics has been across a few different areas, so I wanted to revisit some of the fundamentals I learned a few weeks ago: How robots move.

Say, for example, we have a quadruped like in the video (huge shoutout to my CS123 class), movement comes down to a few core concepts, which can be applied to any independent kinematic chain.

1) Actuators and PD Control

Actuators (typically electric motors) generate the torque that moves a robot's joints.

One common way to control these joints is through PD (Proportional Derivative) control.

τ = Kp(θ_target − θ) − Kdθ̇

Kp drives the joint toward its target angle, while Kd provides damping to reduce overshoot and oscillation.

By continuously measuring joint position and velocity, the controller adjusts torque to keep the joint tracking its target.

2) Forward Kinematics (FK)

Given a robot's joint angles and physical geometry, FK calculates the position and orientation of its end effector, such as a foot or hand.

The math comes down to coordinate transformations. Each joint rotates or translates its own coordinate frame relative to the previous one. We represent these transformations using 4x4 matrices, which combine rotation and translation.

By multiplying these matrices together, we can calculate the foot's position relative to the robot's body. In Python, this can be implemented with NumPy by constructing a transformation matrix for each joint using its angle and the robot's link dimensions, then multiplying them in sequence.

(In the video, I'm physically moving Pupper's leg while the simulation uses these calculations to predict its position in real time from the measured joint angles.)

3) Connecting Everything

Forward kinematics tells us where the foot is for a given set of joint angles. Inverse kinematics (IK) solves the reverse problem, finding joint angles that place the foot at a desired position. This is something I will revisit tomorrow.

Regardless, the full process looks like - Target foot position to IK to target joint angles to PD control to actuators to physical movement.

My biggest takeaway is how these foundational ideas fit together. Kinematics connects joint angles to positions in space, while control turns desired joint motion into physical movement.

Even as robots become more intelligent, learned policies can still rely on these classical robotics foundations to execute their actions.

[Aarav Kothari @aarav_kothari_x]

This is Day 1.

From now till the end of 2026, I am going all in on robot intelligence.

I’m 18, a freshman at Stanford, and have been obsessed with this space. I will be sharing everything and anything, from things I am learning to new contributions and builds.

With how quickly research in this field is moving, there's always something new to explore. Excited to learn from and connect with others working in this space.

[Aarav Kothari @aarav_kothari_x]

@praneethsamine @dreamlabs_ai YAM arms are goated

[Aarav Kothari @aarav_kothari_x]

Just attended the first @a16z speedrun hackathon. Met some cool ass people and learned a lot, will share more soon. Thank you for the invite and shoutout @hackersquadio

#speedhacks26

[speedrun @speedrun]

hour 1 of SpeedHacks 💻 sharing more throughout the day...

[Aarav Kothari @aarav_kothari_x]

Insane value

[Ilir Aliu @IlirAliu_]

ETH Zurich just open-sourced their entire 2026 robot learning course.

Not a MOOC. The actual course. Slides, lecture recordings, coding assignments, GitHub repo.

The curriculum goes from imitation learning and RL all the way to Vision-Language-Action models and foundation models for robotics.

Guest lectures from the co-founder of Physical Intelligence. The creator of Diffusion Policy. Pieter Abbeel. Dieter Fox.

12 weeks. Free. No signup.

Taught by Oier Mees and the team at ETH Zurich.

If you want to understand where robot intelligence is actually heading… this is the reading list the field is using right now.

📍[https://t.co/eKsIjILi60]

——

Weekly robotics and AI insights.
Subscribe free: https://t.co/9Nm01QUcw3

[Aarav Kothari @aarav_kothari_x]

Despite decent exposure, revisiting the foundational work behind modern robot learning is needed and overdue

I'm starting with Zhao, Kumar, Levine & Chelsea Finn’s ACT paper where something simple in hindsight became incredibly influential

https://t.co/53SzgpNtdW

[Aarav Kothari @aarav_kothari_x]

Recently built a simple behavior cloning policy on LIBERO w/ 50 demos

Adding proprioception didn’t fix initial failures. Instead, an issue was the gripper (binary action treated like a continuous one)

Got task solved at step 156 with separate gripper head

Digging into ACT next

[Aarav Kothari @aarav_kothari_x]

Thought I'd start voicing my thoughts on physical intelligence

WAMs replacing VLAs

Generating future world states gives a richer physical prior but adds latency at inference

Is test time world modeling the goal or will it become a training objective for fast action policies?

[Aarav Kothari @aarav_kothari_x]

@sidd_dhanasekar @speedrun @SuperteamUSA Goated

[Aarav Kothari @aarav_kothari_x]

Very good paper

[Praneeth Samineni @praneethsamine]

A few takeaways I had after reading the ROSA paper:

1) Robotics foundation model serving is still primarily treated as an edge computing problem rather than offloading inference to a shared GPU server.

2) This applies well outside of defense, in places like factories and restaurants, where packet loss and unreliable connectivity are less of a concern.

3) Also the general trend is that we seem to be focusing on optimizing a single action model, rather than designing the system around everything else a robot may need, including safety checks, high-level planning, and task progression models.

Definitely recommend giving it a read.

https://t.co/g7eklhQEbS

[Aarav Kothari @aarav_kothari_x]

@ClaraChengGo ☠️

[Aarav Kothari @aarav_kothari_x]

Any must reads after these?

- Walter Isaacson’s elon musk and steve jobs
- Homo Sapiens
- The Power of Now
- The Almanack of Naval Ravikant
- The Mom Test
- Atomic Habits
- Dan Brown series + Wool series

[Aarav Kothari @aarav_kothari_x]

Looking into the coordination + memory layer for robotic fleets (heterogeneous and homogeneous). Seems extremely promising

[Aarav Kothari @aarav_kothari_x]

Graphify — Open-Source Knowledge Graph Skill for AI Coding Assistants https://t.co/pZgX8Vo40G