Physical Intelligence

“Image tokens contain the scene layout in early layers and the target noun carries which object to pick. In deeper layers, the target's location emerges in text and readout position”

“all four variables are recovered in every model, with cross-model Procrustes similarities well above a shuffled-variable null”

“the vision-language backbone does not need to pass information to the action expert at every layer. Instead, there is a 'narrow middle-to-late region through which the backbone conditions action generation'”

“five π0.5 checkpoints were evaluated inside Ropedia W0's proxy environment with r = 0.984 ranking correlation”

“Sergey Levine... A leading figure in reinforcement learning and robotics”

“Physical Intelligence (Pi) cited as a research-oriented robotics lab”

“Kevin Black... Co-author of the π0.5 paper [Ref 12]”

“We use GPT-6-Astra as System 2 and π0.5 [12] fine-tuned on RoboDojo as System 1 (VLA)”

“A single, general-purpose model (GPT-6-Astra) can outperform specialized Vision-Language-Action (VLA) models across a wide range of robotic tasks without updating its weights.”

“Table 9 shows that task memory constructed by four different frozen VLMs”

“Chelsea Finn... A prominent researcher in meta-learning and robotics”

“Black is a key author of π0 and π0.5 at Physical Intelligence”

“Contrary to prior belief, discrete action parametrizations, even discrete regression (DR) with the simplest bin discretization, can exhibit greater control precision than their continuous counterparts, despite the possibly lossy quantization of actions!”

“discrete flow is competitive with, and often slightly surpasses, continuous flow. More surprisingly, discrete regression consistently outperforms continuous regression.”

“On single-subgoal tasks, the low-level policy achieves 85.4% success in simulation and 66.7% on real WidowX hardware. But performance drops to 26.9% at 2 subgoals and 0.0% at 5 subgoals in simulation.”

“π0, SmolVLA, and GR00T employ diffusion or flow-matching action experts”

“It is pairing its robot bodies with AI from Physical Intelligence rather than building the full stack alone.”

“The π0.5 variant also closes the laptop and places the dumpling in every trial, but it drops the coat after grasping it in four trials”

“GR00T N1.7 and Ψ0 reach 16.7% and 0%, respectively for locomotion; GR00T N1.7 completes 0/35 trials and Ψ0 completes 3/35”

“state-of-the-art VLAs such as pi_0.5”

“We show that there exists an effective and efficient complementary approach: the right reasoning recipe can substantially improve the zero-shot task performance of existing state-of-the-art RFMs.”

“We use external generation because the built-in language components are less reliable at generating these traces than at using them for control”

“On a Unitree G1 humanoid, Long-WAM achieves 90–100% grasping success across conveyor speeds from 3.0 to 7.5 cm/s”

“Pi0.5 fails completely on dynamic cup stacking (0/20 trials) and achieves only 15% grasping success at 3 cm/s conveyor speed”

“VLA policies inherit the long-tail coverage problems of their text-action training data, while goal-image planning sidesteps language entirely.”

“EgoLAP introduces a method to train robots using large-scale egocentric human video data, bypassing the expensive process of collecting robot-specific demonstrations.”

“Following π0.5, the normalized state is discretized into 256 uniform bins and appended to the prompt as text”

“Most leading VLA companies (Physical Intelligence with π0, OpenVLA, CogACT) use this direct mapping.”

“robotic planning [Black et al., 2024] have seen successful implementations of generative models using flow matching”

“wrist cameras, which are used by nearly all existing bimanual manipulation systems [3, 4, 5, 6]”

“GT-VLA: Target-Conditioned Trace Guidance for Generalizable Robotic Manipulation”

“Qwen 3.8 Max was tested in the VLM swap experiment (34.2% avg success). Qwen3-VL-4B-Instruct was used as the base for the LoHo-Manip reproduction's fine-tuned manager.”

“Even π0.5 — a state-of-the-art VLA from Physical Intelligence/Google — scored 0.0% on LIBERO-Pro across all six settings”

“GPT-5.6 Terra was swapped in at inference time without retraining, achieving 37.5% average success — competitive with the default Gemini configuration.”

“Vision-language-action policies form a complementary set of output interfaces over various tokenizations of action”

“GT-VLA's core innovation is using an off-the-shelf generalist VLM (no fine-tuning required) to decompose tasks into skills and identify semantic target points, then converting those points into 2D image-space end-effector traces that condition a separate action policy.”

“Gemini 3.1 Pro is used for both SEAL baseline and GT-VLA's semantic guidance. Gemini 3.6 Flash is tested in the VLM swap experiment.”

“Physical Intelligence is a leading competitor in VLA models. Rho directly benchmarks against π0.5 and outperforms it on physical robot tasks.”

“Co-author on RoboMME [5], DIRECT [6], MemER [20], π0.5 [13], π0.7 [14], and Hi Robot [19]”

“DexAgent takes one egocentric RGB video of a human performing a task and converts it into 500 physically verified robot training episodes with three camera views”

“"a single demonstrated task can be expanded into a larger set of physically feasible training episodes with varied object placements and viewpoints" (Appendix A)”

“Perry Dong, a Stanford researcher, and Chelsea Finn... have written a nice piece about what is holding back robotics.”

“LEDGER: a SAM3 tracker maintains per-object stationary intervals and occlusion events, a VLM generates demonstration transcripts, and an LLM planner reads the record at step boundaries to decide what the policy should act on.”

“Robotics is sitting almost exactly where language modeling was: The pretraining has scaled beautifully. What is missing is the other half: the model learning from its own experience, and for that, we need a recipe for post-training.”