Physical AI

Proxy Policy Steering: Adapting a Frozen Generalist Policy at Inference

Proxy Policy Steering trains a reference proxy and a task proxy whose calibrated velocity-space residual steers a frozen base sampler at every denoising step. Across 8 real-world and 4 simulation tasks it lifts the pi-0.5 base by 53% absolute success on average while keeping broad capabilities such as failure recovery.

Assembling Two Parts in One Hand

In-hand assembly mates two rigid parts inside a single 22-DoF dexterous hand, with one group of fingers holding and stabilizing while another aligns and inserts. Trained purely in simulation with RL, the policies transfer zero-shot to hardware using one camera; closed-loop assembly reaches 15/20 (bottle), 17/20 (syringe) and 16/20 (marker) against 2/20, 0/20 and 0/20 for open-loop replay.

Carveout Brings Object-Level Understanding to 3D Gaussian Splat Scenes

Carveout propagates segmentation from rendered views back down to individual Gaussians to give existing 3D Gaussian Splat scenes object-level understanding: it renders views, segments and identifies objects, accumulates evidence across views, then groups labeled Gaussians into 3D instances with bounding boxes and metadata. The current version runs fully locally on SAM 3 and Qwen 3.8 27B, with a reworked web UI, and is due for release in the next few days.

SimWorld Studio: Evolving Environments for Embodied Agents

An Unreal Engine 5 platform where coding agents build verifiable training worlds that co-evolve with embodied learners.

mjbatch Runs Thousands of MuJoCo Sims in Parallel

Thousands of MuJoCo sims run on CPU in parallel for RL, MPC, SysID, and design.

CAP: Continuously Adaptive Perception-Blind Humanoid Locomotion via Learned Denoising

Humanoid locomotion across complex terrain demands forward-looking exteroception to anticipate obstacles, yet this signal is unreliable in real-world deployment, failing partially and intermittently. Existing perceptive policies often assume that depth observations remain clean and in-distribution, while recent attempts to unify perceptive and blind control typically route or switch between separate sub-policies, leaving recoverable information in partially corrupted depth unexploited. We instead propose CAP, a single-stage humanoid locomotion policy that recovers this signal with a perceptive world-model encoder trained as a learned denoiser to reconstruct clean depth from a corrupted input, together with a co-active proprioceptive variational encoder that supplies depth-free body-state information. A coupled training recipe pairs a depth-noise curriculum on the world-model input with world-model feature dropout on the policy-facing latent, exposing the policy to failures across the entire perception-quality spectrum. In simulation, CAP matches or improves upon perceptive baselines when depth remains informative, and degrades more smoothly than a binary-switching baseline as perception worsens. On the Unitree G1, controlled trials and indoor-outdoor deployments demonstrate perception-robust locomotion under intermittent occlusion, real-sensor corruption, and outdoor depth artifacts.

Chen, Hongjin, Xu, Zijun, Ma, Shihao

Show-Harness: Just a VLM Agent Can Play Robots

Foundation vision-language models (VLMs) exhibit broad intelligence about the world, yet translating this intelligence into robot control remains challenging. We present Show-Harness, an Embodied Harness that enables VLMs to "play" robots through a compact semantic interface linking intent to action. Show-Harness exposes discrete semantic action units that VLMs can naturally reason over, while embodiment-specific interpreters deterministically ground them into local robot actions, keeping the VLM directly responsible for fine-grained physical decisions. Through the same interface, Show-Harness demonstrates the feasibility of (1) directly unlocking closed-source frontier VLMs for zero-shot robot control, and (2) adapting small-scale open-source VLMs for low-cost deployment with just a few GPU-hours of fine-tuning. We further develop GUMI (GUI Manipulation Interface), which extends the same semantic action space to GUI-based demonstration collection, allowing humans and agents to "play" robots across embodiments without specialized teleoperation hardware. Extensive experiments show that Show-Harness-equipped VLM agents generalize robustly across tasks, embodiments, and environments, outperforming representative agentic and VLA paradigms. These results suggest that the right interface can unlock substantial embodied capability from foundation VLMs, without requiring additional model capacity or costly embodiment-specific pretraining.

Yanzhe Chen, Zechen Bai, Zhijun Cao

Fire3D Turns RGB Video into Simulation-Ready 3D Scenes

Fire3D converts an unsegmented image or casual RGB video into a complete textured, simulation-ready 3D scene in under a minute. It discovers objects and estimates their 6-DoF poses without supplied masks, boxes or prompts.

StereoPolicy Learns 3D Structure Directly From Stereo Images

StereoPolicy removes the need for depth sensors, depth maps, point clouds, and LiDAR by learning 3D structure directly from stereo image pairs. Across five real-world tabletop manipulation tasks, it reaches 59% success compared with 42% for RGB, 41% for RGB-D, and 14% for PointNet.

Bimo: An Open-Source Biped You Can Run From a PC, an SBC or Its Own RP2040

Bimo, an open-source bipedal robot, can be driven from a PC, fitted with an SBC for onboard compute, or run its walking policy directly on the RP2040. The repo covers API, IsaacLab training, MCU firmware and ROS2.

GPT-6 Astra Makes a Unitree Go1 Climb Stairs Blind

GPT-6 Astra drives a Unitree Go1 up stairs in simulation: instead of writing gait code it emits joint-angle sequences directly, so this is blind locomotion with physics paused between inference calls. The author ties it back to the 2023 GPT-4 trajectory-generation work.

GPT-6 Astra Drives a Robot Arm From a Human Video on the First Pass

Drop a recording of a human doing a novel task into the codex app, prompt it to drive a robot arm the same way, and GPT-6 Astra works on the first pass: physical in-context learning out of the box.