BLOG

Blog

Empirical Results Fine-Tuning pi0.5 on a Real Manufacturing Task

Dream Machines fine-tunes pi0.5 on a bimanual actuator unboxing task: 76% with 21h of data, 90% after one clean hour, 93% from RTC serving settings alone, while LoRA and training from scratch fall far short.

VLAFine-tuningPi0.5

Dream Machines· 2026-09-23T00:00:00

RLT: Physical Intelligence masters the last millimeter with RL Tokens

Physical Intelligence (Sergey Levine's team, Mar 19, 2026) present RL Tokens (RLT): freeze a pretrained VLA, attach an encoder-decoder that compresses its internal embeddings into a bottleneck RL token, then run online RL with a ~1M-parameter actor-critic on the real robot, refining only the critical phase (editing VLA action chunks rather than generating from scratch, anchored by a BC regularizer plus reference-action dropout). Across four sub-millimeter tasks the critical phase speeds up by up to 3x, screw success goes 20% to 65%, and half of Ethernet insertion episodes beat every human teleoperation demo - with just 15 minutes of real robot data.

VLAOnline RL

Physical Intelligence · 911 views · Mar 19, 2026

reBot-DevArm Deep Read: 96 STEP Files, a Screw-Level BOM and an Overnight VLA Loop

Seeed's open-source six-axis arm in full: 96 STEP files, a screw-level BOM recomputed line by line (motors = 89.9%), the published thermal-test limits, the CERN-OHL-W-2.0 licence switch, and the LeRobot + GR00T N1.7 overnight autonomy loop.

Robotic ArmSeeed Studio

铁卫士 / RobotWorld 整理 · 1.7k views · Sep 20, 2026

Zetta ζ: Closed-Loop Self-Evolution for a Frozen VLA Policy

Tsinghua AIR and Z-Trans AI present Zetta ζ, a closed-loop embodied harness that improves a frozen VLA policy without a single gradient update. Instead of fine-tuning, it evolves the execution harness around the policy: code-based runtime critics watch every action, recovery skills take over on deviation, and a validation gate admits only skills that generalize. The frozen baseline scores 31.0% on LIBERO-Pro; the same policy under Zetta ζ reaches 92.5% (+56.3 absolute points), with a +20-point gain to 93.6% across 18 RoboCasa tasks. Z-Infra scales valid rollout throughput 20.6× and speeds inference 11.1×. Skills transfer zero-shot, and clear robotic "aha moments" emerge.

Embodied AIVLA

Tsinghua AIR & Z-Trans AI · 936 views · Aug 30, 2026

Perceptron Open-Sources Isaac 0.5: First Open Model at the Frontier of Video Understanding, Embodied Reasoning and Robot Control — 210x Less Teleop via Video Scaling

Isaac 0.5 is Perceptron's open-source embodied foundation model with 36B sparse parameters: it reads images, video, language, robot state and previous actions to answer video questions, point and track objects, report task progress, and generate robot actions. The team establishes a scaling law trading video for teleop: scaling general video from 1,000 to 1M hours cuts the teleoperation needed for action loss 2.50 from ~5,900 hours to 28 (210x). Trained on 35+ robot systems, 100K hours of robot experience, 1M hours of video and 3T multimodal tokens, it introduces semantic world modeling (predicting future percepts), the mHarmony typed multimodal interface, and Null Experts for dynamic compute — leading all five perception task families at 8.5x lower inference cost. Weights, training code and LeRobot inference code are fully released.

Embodied Foundation ModelScaling Law

Perceptron · 830 views · Aug 26, 2026

Introducing S1: In-Context Learning for Robotics — One Video Demonstration Is All It Takes

Skild AI's flagship robotic foundation model S1 is built from the ground up as an in-context learner: show it a video demonstration of a task — even an unseen 10-minute long-horizon one — and it executes with no fine-tuning. On unseen tasks, one demonstration ≈ 380 post-training examples; at 100k pre-training hours S1 reaches 66% success vs 9% for language-prompted VLAs.

Skild AIS1

Skild AI · 909 views · Aug 18, 2026

Helix Learns to Fold Laundry: First Humanoid with Multi-Fingered Autonomous Folding

Figure AI's Helix VLA model achieves first humanoid multi-fingered autonomous laundry folding. Same architecture seamlessly transitions from logistics to household chores with data-only change.

Figure AIHelix

Figure AI · 662 views · Aug 25, 2026

Helix 02: Figure's First Unified Whole-Body Loco-Manipulation VLA

Figure AI releases Helix 02, extending single-network control from upper body to whole robot. System 0 trained on 1000+ hours of human motion data replaces 100K lines of hand-engineered C++, enabling 4-minute continuous autonomous loco-manipulation.

Figure AIHelix

Figure AI · 697 views · Aug 25, 2026

The Developer Inflection Point: Leju's Open-Source LeTools Post-Training Toolchain Compresses Embodied-AI Development from Weeks to Days

Across ICRA REAL-I and CRAIC 2026, 200+ student teams used Leju's open-source LeTools chain to go from algorithm to real-robot deployment in as little as one day. Hardware-free dry-run of a 37-node behavior tree, a 1,000-hour production-grade LET-Base dataset, a one-line simulation-to-real switch in deploy.yaml, and 10+ architectures behind a unified Adapter layer show embodied-AI competition shifting toward low-barrier dev environments covering data, models, and deployment.

LeToolsSim-to-Real

深蓝AI · 601 views · Aug 15, 2026

LingBot-VLA 2.0 Reproduction Radar: Weights Open-Sourced, but 60K-Hour Pre-Training Cannot Be Equivalently Reproduced

Official 6B base weights, RoboTwin post-training weights, training code, open-loop eval, simulation eval, and deployment entry are all public. Labs can run inference, RoboTwin closed-loop evaluation, or custom-data post-training, but the full 60K-hour raw corpus and training budget are missing, making equivalent base pre-training reproduction infeasible.

LingBot-VLAVLA

TPAMIs · 647 views · Aug 6, 2026

Two Hands Are Harder Than One Brain: Interpreting the VLA Bimanual Manipulation Survey

A 35-page survey by Chef Robotics' Inkyu Sa covering 200+ sources and 31 methods from a production-floor perspective. It organizes bimanual VLA research by 'coupling tightness × action-head emission,' argues flow-matching heads with action chunking are the strongest dual-arm recipe (winning on ~10 integration steps), and dissects the two-orders-of-magnitude gap between benchmarks and real deployment plus three durable gaps (force, evaluation, data).

VLABimanual

Inkyu Sa (Chef Robotics) · 631 views · Aug 9, 2026

As an Amazon Associate, we earn from qualifying purchases.