Léo

5 months ago, we started hosting the @roboticsstack podcast. 23 episodes in, here are some lessons on what Robotics businesses actually work: - boxing shoes is still an open problem - teleop arbitrage: bill $30 an hour, pay an operator $11 - sell the hardware, let the customer find the customer - complexity grows exponentially with the number of integrations - your buyer is one overwhelmed IT guy 🧵👇

NEW DEXTERITY LEARNING: Single hand Rubik cube solving! This project involves @YutongLiang_, @QuanquanPeng03, and others. FINGR, for Future-supervised Interaction Network with Geometric Representations, is a learned policy from UC San Diego. It turns the layers of a cube held in one dexterous hand. The hardware is a 22-joint Sharpa hand on an xArm7, with only 13 DoF actually controlled, the thumb and middle finger hold the cube still. Each moving finger carries a 96x96 tactile deformation image. The cube is a 52 mm 2x2x2 carrying coded markers, and external multi-view tracking reconstructs it in the palm frame. In practice, this is a reliable single primitive, with a solver, a tracker, a planner and a table doing everything else. The learned component is one primitive: a discrete solver minimizes positive quarter turns, the system plans pickup, placement, regrasping, state verification and recovery. The neural network does exactly one thing: rotate a layer 90 degrees against contact.

I did a deep dive on a bimanual dexterous rig to solve a Rubik's cube, trained on a CPU only, and open sourced the whole project. Main number: 122 of 128 scrambles of 20 moves solved -> 95% success rate. A median solve is 54 commands plus 25 reseats, 8.4 simulated minutes. The setup: - two Allegro hands in MuJoCo 3.14. - left hand holds the cube palm up and the right hand works from above. - the cube is three bodies on two hinges, with friction. - benchmarked on 20 moves because that's pretty much the baseline to solve any Rubik's cube (afaik). - Kociemba's two-phase algorithm solves the cube from each of its 24 orientations. - each solution is translated into commands the hands can execute, and I take the most direct one. The hands know five commands: U, U', D, D' and a tip that rolls the whole cube 90 degrees. Each command is learned in three steps: 1. a scripted teacher with feedback performs it 2. an MLP (512-512-256, rsl_rl) clones the teacher from noisy demonstrations: 240 inputs, 13 coordinated controls, 50 Hz 3. DAgger rounds, where the policy drives and the teacher labels the states it reaches Each training run took 1 to 3 hours on 10 CPU threads. What made it work: - Coordinated controls: imitating 38 separate servos failed, small errors in one hand shook the cube loose. - No PPO: fine-tuning made the single-move policy worse, so I dropped it. - A reseat: each tip leaves the cube about 2 degrees off. Squaring it up before a command took the same 32 scrambles from 2 solved to 32. - Recovery: up to 3 per run, each re-planned from where the cube is. Before this, I tried spinning a pen with one Sharpa Wave hand, and it never got past two revolutions. RL gave one clean turn on 194 of 200 held-out starts, but the second revolution only counted once the grip was gone, and no full spin and return cycle ever certified. -> imho the cube worked because a planner splits it into short commands a teacher can script. A pen spin is one continuous finger gait, and PPO, cloning an MPC teacher and DAgger all stalled on it. The repo has the code, trained policies, start banks and training commands under Apache 2.0, and one command reproduces 32 solves in about 15 minutes on 10 CPU threads. - >If you want to build on this, work on dexterity together, or anything else, my DMs are open.

NEW HUMANOID CONTROL: Humanoids can now catch an object without falling! Called Reflex, for Visually Grounded Reactive Humanoid Control, it is a system from the University of Washington, USC, Stanford, NUS, UNC Chapel Hill and AI2. It makes a Unitree G1 catch boxes thrown at it by people, using only an onboard RGB-D camera. The robot does not catch with hands: it repositions its feet and torso, forms a support with its forearms and chest, absorbs the momentum and keeps hold without falling. Training runs in three stages in Isaac Lab 1. train a state-based teacher with PPO plus an AMP-style reward from human motion references 2. training with PPO while learning to infer the box's dynamics from delayed and incomplete observations of position, velocity, orientation, size and observation age, compressing them into a latent that conditions the controller 3. freeze the controller and learns to recover that same latent from 25 frames of egocentric RGB-D history plus camera ego-motion, supervised with latent matching, an auxiliary box-state loss, heatmap supervision and DAgger. Since rendering never enters the reinforcement-learning loop, visual training needs "only" about 1.6 million rendered steps. I find it interesting that the policy driven by camera images beats the policy handed perfect simulator state. Here the student's interface, a learned latent rather than raw state, plus more RL through it, produced a better controller than ground truth.

NEW BENCHMARK: Coding agents write the robot simulations! Gym-Anything for Robotics is a pipeline from Carnegie Mellon in which coding agents write robotics simulations instead of humans building them. It runs in four phases: 1. the choice of what to simulate is grounded in US labour data: 1,016 occupations and 465 everyday activities are decomposed into 31,961 atomic physical actions, weighted by wage or time spent, grouped into 2,330 action groups, and an optimization picks 120 that cover the most valuable physical work 2. a creation agent writes each simulation as code, scene, physics, robot and rewards, and produces evidence that it is good, which four audit agents including an adversary check against ten criteria over three rounds 3. a task creation agent writes eight tasks per simulation, from entry level up to one test task, each checked by six more audit agents and each scored by a verifier reading the simulation's physical state 4. then frontier models are studied learning in the result The output is RoboWorld: 25 built simulations covering liquids, cloth, chemistry, molten metal, soil, soft tissue, pastes and acoustics, with 25 test and 175 training RL environments, swappable robots from single FR3 arms to Unitree G1 and H1-2 humanoids, and 10,000 variation seeds each. -> The pipeline, environments and benchmark are released!

Dreame made robot vacuums. It just raised $100M to build humanoids. Xiaomi made phones, now it ships cars. Chinese companies see their strength as the ability to adapt to whatever is needed next. Volkswagen would never. On @roboticsstack. Full episode 👇