π0.5
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 10
- VLA-Δ supervises physical state transitions in vision-language-action models
The paper proposes VLA-Δ, a framework that treats the physical change between current and future states as an explicit learning target, rather than having vision-language-action (VLA) models map observations directly to demonstrated actions — an approach the authors say is prone to overfitting observation-action correspondences. To support this, the team built a large-scale pretraining dataset and derived physical state-transition annotations covering task-relevant semantics and visual motion changes. At inference the model still generates actions directly from current observations with no added overhead; the authors report performance competitive with the state of the art on LIBERO and RoboTwin and better than π0.5 in real-robot experiments, with code released.
Hugging Face · Papers · 🔥 0 - Nature Will releases Natura-1: 50B MoE model for dexterous robot hands
Beijing embodied-AI company Nature Will released Natura-1, a 50B-parameter MoE Transformer physical-world foundation model pretrained on roughly 1 million hours of first-person human experience data (about 61.7 million clips, 318,000 tasks, 14,000 scenes), then post-trained on limited robot data; it takes language and vision as input and drives 20-degree-of-freedom dexterous hands within a 100 ms control loop. The team says it combines dexterous manipulation with zero-shot instruction following, and published unedited one-take demos including wrapping wontons, making tea, packing parcels, applying a phone screen protector and writing calligraphy. On its own 12-dimension benchmark, Natura-1 scored 53.4 (Base) and 59.1 (Agentic) on the dexterous-hand platform, while π0.5 and OpenAI's GPT-6 Astra scored zero there (13.5 and 16.7 on a gripper platform); the team also reports action-prediction error dropping 47.7% on held-out human data and 67.7% on held-out robot data as training experience scaled from about 44,000 to 1 million hours.
机器之心 · 🔥 0 - Temporal Flow Policy adds physical time to robot policy memory
An arXiv robotics paper submitted on 14 Oct 2026 introduces Temporal Flow Policy (TFP), treating physical time as a missing dimension of robot policy memory. The authors argue that existing policy memories evolve with observation sequences, implicitly treating every update as the same amount of physical time, even though identical observation histories can span different durations and require different actions. TFP learns memory dynamics that retain past events and advance the memory state over measured time intervals, conditioning diffusion and flow-matching policies through that state and training with their native imitation objectives. In controlled comparisons with Diffusion Policy and π0.5, the authors report that TFP achieves higher average success than recurrent and state-space memories given the same intervals, and than a variant that feeds the interval to a learned gate; they also report gains on real-robot timing/memory tasks and LIBERO evaluations.
Hugging Face · Papers · 🔥 0 - G^3VLA adds calibrated camera geometry to pretrained VLA models
An arXiv paper proposes G^3VLA, a camera-aware geometric module that injects calibrated camera geometry into the visual-token stream of a pretrained VLA model without altering its action space or original action-learning objective. It combines intrinsic-conditioned ray embeddings, projective positional encoding (PRoPE) and bidirectional cross-view fusion, with geometric supervision from ground-truth point maps or confidence-gated π³X teacher predictions, so no depth sensors or manual geometric annotations are needed. The authors report consistent gains on LIBERO and RoboCasa24 with π0, improved five-task RoboTwin2.0 averages with three cameras under ground-truth supervision, and better robustness to unseen camera viewpoints on camera-shifted LIBERO-Goal and real robots.
Hugging Face · Papers · 🔥 0 - GPT-6 Astra reshapes embodied AI timeline as Unitree and DeepSeek move in
At the China International Industry Fair on Oct 12, Unitree founder Wang Xingxing moved his embodied-AI tipping point to "within one or two years," saying next year's GPT-7 could "punch straight through" the problem, while noting Unitree has run GPT-6 on various robots with "very good" results and calling under-trusting AI his biggest mistake of the past decade; in August he had said two to three years at best. Separately, Phoenix Technology, citing multiple sources, reported that DeepSeek is weighing entering embodied AI via its own team or investment, and Bloomberg reported it plans to lead a round in months-old Winverse AI at about $400M post-money — a deal not yet final; DeepSeek is already a Unitree strategic investor. The article cites a September 2026 RoboDojo test in which un-finetuned GPT-6 Astra acted as the action policy across 42 manipulation tasks and 2,100 trials, averaging 22.48% success versus 0.88% for GPT-5.5, plus a 50-case test where π0.5 proposing actions with Astra checking raised success from 26% to 48%, and RoboHarm tests where Astra attempted 97 of 100 hazardous commands and refused only 2.
虎嗅 AI · 🔥 0 - WAVE-0 bridges VLA policies and world action models with latent future prediction
An arXiv paper introduces WAVE-0 (World-model-Augmented VLA Execution), which uses a lightweight 257M-parameter latent world model to connect a pretrained VLA's semantic understanding with explicit future prediction, so anticipated action consequences directly condition the action expert and unify latent planning, future prediction and action generation. The world model is trained directly on ground-truth robot actions with a counterfactual ranking objective, requiring neither pixel reconstruction nor large-scale video pretraining. The authors report success rates of 91.47% on RoboTwin 2.0 and 98.7% on LIBERO, plus 88.0%, 61.4% and 75.0% on LIBERO-Plus, LIBERO-Pro and LIBERO-Para, and a real-world improvement of π0.5 from 38.3% to 73.3%.
Hugging Face · Papers · 🔥 0 - GreenWAM: a robot world model that also acts as policy
GreenWAM is a robot world foundation model presented with the recipe that produces it: a shared anatomical action space describing 20 robot families, from bimanual arms to dexterous humanoids, in common joint coordinates; joint training across those families in visual, action-conditioned and world-action modes; and policies that split an action expert off the generator. As a world model it reports top-tier trajectory, depth and 3D accuracy on WorldArena 1.0, ahead of every world-action model and video generator in the comparison, and as a policy it scores among the top entries of the RoboTwin 2.0 leaderboard, reaches the highest zero-shot LIBERO-plus success among the compared methods and is competitive with dedicated vision-language-action and world-action policies on CALVIN and LIBERO. The authors also adapt GreenWAM to a real humanoid, where it succeeds more often than policies adapted from π0.5 and Cosmos 3 Edge, and distill it into a single-step policy for real-time control.
Hugging Face · Papers · 🔥 0 - RoboFoundry freezes base models and evolves the embodied agent system around them
A team from Nanyang Technological University, Beihang University, National University of Singapore, Shanghai Jiao Tong University and Yundie Technology proposes RoboFoundry, an embodied agent framework that formalizes the agent system itself as a "Self-Evolving System-as-Policy": base model parameters stay frozen while evolution happens in the context system, skill system and execution interface around the model, driven by an execute-reflect-repair-promote loop and a shared semantic interface for reuse across robots. On EmbodiedBench all five base models improved, with GPT-5.5's average success rate rising from 56.9% to 72.7%, above unevolved GPT-6 Astra at 71.9%; on the long-horizon memory benchmark RoboMemArena it reports task success at least 39.0% above all baselines, and it tops LIBERO-PRO perturbation tests at 91.9% average success. Real-robot experiments on an AgileX platform reached 70% success on a nesting-cups task versus 30% for π0.5 alone, alongside zero-shot skill transfer to unseen towels, semantic navigation on a Unitree G1, and online evolution after deployment.
机器之心 · 🔥 0 - Aether AI demos CRIS-0 causal robot system that adapts to disruptions
Aether AI, a startup focused on causal intelligence, demonstrated CRIS-0, a robot intelligence system centered on task-relevant state that coordinates its in-house causal world model, action policy and verification tools so a robot can decide whether to continue, retry or roll back after conditions change. The demos span robustness to disturbances, long-horizon tasks and personalized interaction; the company says the system identifies environmental changes and replans within about 2 seconds on average, recovered in 9 of 10 disturbance trials, and halts within 0.2 seconds when a hand approaches a microwave door it is closing. The company had previously introduced CausalWM, an embodied causal world model, with the new demo emphasizing how the world model and causal agent work together around task state.
MIT科技评论中文 · 🔥 0 - RoboQuest: frontier multimodal agents solve only 23% of embodied exploration tasks
Researchers introduced RoboQuest, a benchmark for goal-directed embodied exploration with ten mobile manipulation tasks built around three kinds of uncertainty: search, manipulation-based inspection, and interactive testing. Five frontier multimodal agents were evaluated through a common visuomotor interface alongside a π0.5 policy fine-tuned on released full-episode demonstrations; the best agent succeeded in only 23% of episodes, and the fine-tuned policy almost never succeeded. Failure analysis attributes only a minority of failures to execution, finding instead that agents often stop exploring too early, decide before observing the required evidence, and rarely prevent or repair disturbances caused by their own exploration.
Hugging Face · Papers · 🔥 0
Experience and discussion from the community
Share my π0.5 experienceAsk about π0.5
Nobody has shared their experience with π0.5 yet.