Tao

Building self-driving race cars (A2RL) & soccer-playing humanoids (RoboCup) @KhalifaUni. Ex-C++/Android dev. Vim & Spacemacs user. Corgi dad.

[Tao @TaoBuildsRobots]

1/
I build autonomous race cars and teach soccer-playing humanoid robots to walk.

Researcher at Khalifa University, Abu Dhabi.

Here's what I work on, and what you'll see on this account:

[Tao @TaoBuildsRobots]

@chelseabfinn Curious whether this helps on locomotion-style problems too, where rewards are dense and policies are small, or if the gains mostly show up with large pre-trained policies.

[Tao @TaoBuildsRobots]

@stisov16 Pipetting with a hobby arm at a hackathon is a great demo. How are you handling repeatability? Hitting the same well every time seems like the hard part.

[Tao @TaoBuildsRobots]

Question for robotics people: what's your best "worked in sim, fell over on the real robot" story?

Mine: a walking policy that looked great in MuJoCo. On the real robot it managed a few steps, then fell over.

[Tao @TaoBuildsRobots]

@tokufxug Syncing cameras and LiDAR together is the interesting part. My bet for our real race car is that sensor timing breaks first, so I'd love to know if it models latency between sensors, not just what each one sees.

[Tao @TaoBuildsRobots]

@realfxw Ex-Android dev who now trains walking policies here, so this is my two worlds colliding. Curious how much timing jitter they see running the policy on the phone side.

[Tao @TaoBuildsRobots]

@Mascobot Depends on the layer you want to be serious about. We buy our humanoids and put everything into the walking policies. Still feels pretty serious when one falls over.

[Tao @TaoBuildsRobots]

@rdn_nikita Is this a whole-body control model or something higher level? Curious whether teams with smaller humanoids will be able to fine-tune it.

[Tao @TaoBuildsRobots]

@LeRobotHF @microagi @huggingface Millions of hours of human data is great for manipulation. Curious if anyone's using it for humanoid locomotion yet, or if walking still mostly comes from sim.

[Tao @TaoBuildsRobots]

My corgi has better balance than our robot.

Working on it.

[Tao @TaoBuildsRobots]

@dhh Fun bonus for robotics people: 1024 is also a very normal number of parallel environments for RL training. So it's a holiday for the GPUs too.

[Tao @TaoBuildsRobots]

@OpenRoboticsOrg Thanks to the infra team. Most of us pull ROS packages every day without thinking about who keeps those servers patched.

[Tao @TaoBuildsRobots]

@LeRobotHF 600 hours of manipulation data is a lot. Locomotion is the opposite situation: almost no shared real-world data, so most of us still generate ours in sim.

[Tao @TaoBuildsRobots]

@LeRobotHF @Dexmal_AI Is any of the 7 embodiments a humanoid? Curious how far a manipulation-pretrained action expert gets once the base is also moving and balancing.

[Tao @TaoBuildsRobots]

@NVIDIARobotics Multi-physics is the part I'm most curious about. Train a walking policy on one physics backend, test it on another, and see what breaks. Feels like a cheap sim-to-sim check before touching real hardware.

[Tao @TaoBuildsRobots]

Been on Vim keybindings for 10+ years. These days it's Spacemacs, still with Vim keys.

Same muscle memory that wrote Android apps now writes walking policies for humanoid robots.

My fingers haven't noticed the career change.

[Tao @TaoBuildsRobots]

Asked Agility's Jonathan Hurst how they handle robot-to-robot variation across a fleet.

His answer: calibrate what stays fixed, make the policy robust to what changes, and be picky about the second part. Robustness always costs some performance.

[Agility @agilityrobotics]

From @TaoBuildsRobots earlier today: how do you deal with robot-to-robot variation across a fleet? Calibrate every Digit individually, or train policies robust enough not to care?

JH: Both. Minor variations exist in motor constant or robot component mass, that remain consistent once a robot is built, but may vary across the fleet; we believe these can be calibrated. But other features vary with time or temperature or other factors, and for those we must create policies that can handle the random variation. It is important to note, though, that one cannot train a policy to be robust to everything. We, as engineers, have to decide what is likely to vary, and be choosy about what we require the policy to be robust to; mostly, that should be things that will be outside of our control, like ground height, or mass of a bin, or friction coefficients.The more robust a policy is, generally, the lower the performance. For example: If we truly have no idea what the ground friction will be, then the robot always has to walk as if it were walking on ice. If we do have some confidence in the range of uncertainty of the ground friction, we can increase our confidence appropriately for the robot performance.

[Tao @TaoBuildsRobots]

@agilityrobotics Thanks for picking this one up. "Walk as if it's on ice" is the clearest way I've seen that trade-off put.

How do you pick the friction range in practice? Measured at each site, or from experience?

[Tao @TaoBuildsRobots]

@corl_conf @BitRobotNetwork Curious how well head + hand data from people transfers to humanoids whose hands and proportions look nothing like ours. Is there a retargeting step, or is it used more for high-level behavior?

[Tao @TaoBuildsRobots]

@OpenRoboticsOrg Saving these for the weekend. Any talk you'd point to first for someone integrating a full ROS 2 autonomy stack on a race car?

[Tao @TaoBuildsRobots]

@NVIDIARobotics @agilityrobotics Putting compute and safety in the same box is interesting. How does the safety layer sit next to a learned walking policy? Does it watch the policy outputs, or can it take over control?