William Xie

our gripper: github.com/correlllab/h...

but the friction, solimp, solref is unchanged from the 2f85. i noticed greater oscillations prior to implementing that, and breaking up the collision mesh into multiple boxes

It might be helpful to visualize the contact points, forces. Maybe try to debug up your gripper similar to robotiq_2f85/scene.xml in mujoco menagerie, it is a simple box hanging from above. fwiw I set up our lab's gripper just last week in mjc with the same settings as the 2f85 and it works great.

Unsteady Grasp · Issue #786 · google-deepmind/mujoco Hi, I'm a student and I'm trying to use MuJoCo for picking object with UR5 I'm looking for some help with grasping. When I test the pick task in MuJoCo simulator, it seems good and everything is fi...

also of interest: github.com/google-deepm...

If you haven't seen it already, this section of the docs has most of the first recommendations: mujoco.readthedocs.io/en/stable/mo.... Since you are working on a gripper you might want to check out the robotiq 2f85 xml in menagerie, specifically the multiple contact geoms at the finger.

The G-spot refers to the genot-brogliato point in the painleve paradox at which point a slipping planar rigid body lifts-off. from an article in the journal of applied mathematics

🤓🧐🤔🤨

not even thinking about the ethical issues of such labor, which are of course the concerns with primacy.

and sensing!

People are better off splurging for something more proven, if that's within their budget. Spinning up a open/custom hand is more feasible and would actually still be more productive. A lot of people and problems are probably better served by thinking jointly about required morphology and dexterity.

We have had an absolutely horrid time with the Inspire hands and a somehow worse time with the actual H1-2. I don't know what's been improved in the past year but I am extremely skeptical of any hand offering from Unitree, esp for 30k USD.

cam showing up wouldve been nice but not sure that wouldve moved the needle

unexpected brickage from jokic ruining the third, true, and only MPJ we ever needed

yesterday i advised self-censorship to a undergrad student im helping apply for grfp re: their research & personal motivations. i think they've got as good a shot as any other. i think im making the right call, but it feels deeply wrong and capitulatory. any thoughts (esp from reviewers)?

On the Dual-Use Dilemma in Physical Reasoning and Force Humans learn how and when to apply forces in the world via a complex physiological and psychological learning process. Attempting to replicate this in vision-language models (VLMs) presents two challe...

and then to self-plug a bit, i wrote up a little case study/position paper on dual-use in VLM reasoning and robot manipulation last month: arxiv.org/abs/2505.18792. the gist is that safeguarding reduces both helpful/harmful robot control and i opine about what that means for future model eval/dev

seemingly a hot topic rn on my feed but i'm not sure how much more humanity can lower the floor on mass autonomous death, whereas taking care of a human is still wildly inefficient. that is to say on the level of the individual researcher in AI theres a lot more unrealized help we can do than harm

i have two somewhat distinct and existential concerns: 1) that we are guileless researchers operating in aggregate as an arm of the MIC and 2) that our research is directly extensible (within a few degrees) to dual-use. tbh i think many overstate 2) and others are helpless wrt 1) due to "incentives"

Photo of the a large group of GISS employees/colleagues in front of Toms Restaurant with a prominent street sign saying "112th St" taken from the middle of Broadway. Credit: Tricia Baron.

Last day at the GISS building (Wed. May 28th). 🥲

NSF GRFPs terminated for those in graduate school and already attending Harvard, too.

but I think that's a good feeling to lean into

yeah, I think you're right on both points. I got in the weeds on haptic teleop interfaces for LFD recently and overall am not super convinced it'll enable the data scale we need. Way more interested in self-improvement from physical interaction (w/ touch) though I feel quite out-of-depth there

and w/o touch

I'm generally a believer that we'll eventually do everything with vision, but I also believe that we'll need touch to get policies running closer to real-time/humans. My advisor loves to bring up these videos from a study of trying to strike a matchbox w/ and w/o feeling in their fingers:

Me and @wxie.bsky.social

Liebherr

going forward, i'm thinking about how we can scale good data collection with force control and improved physical models & reasoning. as of now you cannot convince me that we do not still need huge amounts of real robot data for robust contact-rich manipulation. and we are quite a ways off...

so interesting where the field has coalesced and where it has diverged. some of it is a necessary byproduct of manipulation, some of it seems like open areas for research. anyway, here's a fun and unreadable plot: these 25 papers evaluate 64 (59 models) significantly different contact-rich tasks

as most data pipelines begin with position-space teleoperated demonstrations, 64% of the reviewed works correspondingly train policies to output end-effector position-space actions

for my area exam last month i reviewed 25 touch-based IL works from 2024/5, taxonomized across tasks, sensor type, model architecture, and more (up now at arxiv.org/abs/2504.11827). here's a perhaps intuitive plot of policy action space--the in-figure citations are clickable in the pdf/html!

I'm giving a lecture on language model debiasing to my undergrad NLP course on Friday but I'm not super up to date on the research. Does anyone have any suggestions for papers/topics to cover?

true, but humans learn implicit control laws, however relative they may be, from rich sensory information over many, many episodes. for robots, high-precision servos are just one tool to obtain such high-fidelity data. i also think such tooling is important for achieving supra-human abilities.

I think the RL policy / teleop comparison here is not quite fair--the RL policy leverages wrench data, which is the primary supervisory signal for these kinds of insertion tasks (learning visuo-force servoing) whereas the teleop here is using a 3D CAD mouse--huge embodiment gap in data collection

taking name suggestions…