Fifteen minutes of a human freely playing with a cube might be worth more than 15 minutes demonstrating the PERFECT rotation. @amazon FAR and @UMBaltimore 's Dexterity from Supervised Play (DSP) first clones goal-free glove teleoperation into a diffusion play policy for a 22-joint hand. It then lets that policy fiddle with objects autonomously, collecting five hours of real contact transitions. A Diffusion Forcing world model learns what each proposed joint-motion chunk would do. When given a target cube orientation, DSP samples 1,024 play motions, predicts their outcomes, executes the best prefix and replans. The goal is introduced at PLANNING time, not baked into the demonstrations. On 128 simulated reorientation trials, it reached within 0.4 radians in 72.7% of cases, versus 25.8% for hindsight behavior cloning. A deterministic play policy covered only 55.4% of tested state transitions; diffusion covered 84.2%, despite similar imitation error. That ablation is the real story. Caveat: real-robot trials use externally tracked cube poses and an RTX 5090 for planning; the paper flags hardware summary as provisional. This isn't onboard general dexterity. My takeaway: teach a robot HOW TO EXPLORE CONTACT first. Teach it what goal matters later. play-policy.github.io/ #RobotLearning #DexterousManipulation

The fastest way to make a robot VLA worse might be to give it MORE 3D information. We keep saying robots need spatial intelligence, so the obvious move is to feed point clouds into a pretrained vision-language model. Except an RSS 2026 paper from @chen_shizhe, Paul Pacaud and @CordeliaSchmid shows how spectacularly that can backfire. PointACT compares WHERE geometry enters the policy. Put point-cloud tokens directly inside a monolithic VLA and RLBench success drops from 73.2% to 18.6%. Yes, adding depth-derived 3D features made the same architecture dramatically worse in that controlled comparison. Their alternative: leave the pretrained Qwen2.5-VL semantic backbone FROZEN and give the action expert its own geometric pathway. A Point Transformer v3 encodes up to 4,096 colored 3D points at five spatial scales. Instead of compressing the entire scene into one geometry embedding, the action tokens visit local windows of point tokens through bottleneck self-attention. Each window updates a copy of the action representation; those copies are pooled, then cross-attend to the VLM's language/image features. Think of it as letting the motor policy repeatedly inspect nearby surfaces while keeping the language model's understanding of the task intact. The ablation is the whole story. On 10 simulated RLBench tasks: Dual-system VLA, no points: 50.8%. Add coarse final-layer 3D features: 69.7%. Naively concatenate multi-scale features: 65.2–65.6%. Actually INTERACT with geometry at multiple scales: 82.3%. More geometry tokens alone weren't the answer. How action tokens interrogate geometry mattered. And they tested physical robots. On an SO-100 arm putting a sock in a drawer, PointACT succeeded 9/10 times versus 2/10 for π0. On a UR5 stacking cups, it scored 7/10 versus 0/10 for π0, although a dedicated 3D policy also scored 7/10. Don't oversell this: those hardware tests are only 10 trials per task, models have different pretraining histories, and depth gets noisy on transparent objects. The authors also show failures where an accurate initial grasp is followed by zero useful recovery. My bigger takeaway: semantic perception and geometric control probably shouldn't compete for the SAME token budget. Let the VLM decide what matters, but let a spatially grounded action expert decide exactly where the fingers go. cshizhe.github.io/projects/p… #RobotLearning #EmbodiedAI #VLA