Skip to content
Only First party#Evaluations and benchmarks
  1. AgiBot72

    AgiBot releases embodied foundation model GO-3 Preview, ranks first on three embodied intelligence benchmarks

    AgiBot has released the embodied foundation model GO-3 Preview, built on its self-developed WRAM (World Reasoning Action Model) architecture that unifies task understanding, future state prediction, and action generation, and it ranks first on three embodied intelligence benchmarks: RoboDojo, RoboTwin, and RoboColiseum. It is planned to be fully open-sourced by the end of 2026.

    Why it matters: The original article fully discloses GO-3 Preview's three-stage design of pretraining, post-training, and post-deployment learning, which can be used to understand the engineering path for scaling embodied models across their full lifecycle.

  1. Dexmate72

    Dexmate and CMU Release MeMoREval Benchmark: Robot Storing Camera Frames Beats Text Descriptions

    Dexmate and CMU introduced the MeMoREval benchmark, which uses a two-stage protocol of experiencing first and asking questions later to evaluate what embodied agents should store as memory. It includes 53 tasks and 4 simulated indoor environments, and runs on the Vega 1 Pro dual-arm mobile manipulator. Results show that storing raw camera frames scored 70% and had lower write cost, surpassing text descriptions at 56 to 60%; adding scene graphs achieved the highest overall success rate among all tested models. The protocol also completed two memory retrieval demonstrations on a real Vega 1 Pro using a SLAM map and head camera.

    Why it matters: The benchmark was proposed jointly by the platform developer and CMU; its frame-versus-scene-graph comparison can help with memory design choices, and the real-robot transfer offers a feasibility reference.

  1. Shanghai AI Laboratory - Embodied Intelligence73

    Shanghai AI Lab launches decision model Intern·Mingjue, outperforming Jev and releasing 0.8B, 2B, and 4B weights

    Shanghai AI Laboratory has launched the decision model Intern·Mingjue (Decision), which outperforms Jev and is available in 0.8B, 2B, and 4B parameter versions. In RTX 4090 tests, the 0.8B and 2B models had an average end-to-end latency of about 33 milliseconds and about 30 decisions per second, while the 4B model was 44.16 milliseconds and Jev was 109.70 milliseconds.

    Why it matters: The original text provides measured latency, benchmark scores, and a mechanism for routing between large and small models, which readers can use to judge the decision model's suitability for interactive scenarios.

  1. PKU EPIC Lab66

    PKU EPIC Lab releases HumanTracker humanoid motion tracking benchmark and HumanScore metric

    PKU EPIC Lab proposed the HumanTracker benchmark for evaluating humanoid motion tracking, accepted to ECCV 2026. The benchmark contains approximately 153 hours of optical motion trajectories from multiple professional performers, organized into four motion families with text labels; the accompanying HumanScore metric is trained on 12K motion pairs containing 24K motions, aligns with human preferences, and can reveal contact and stability failures that kinematic metrics often miss, such as foot skating and mistimed landings.

    Why it matters: The original work proposes a benchmark and metric that align humanoid motion tracking evaluation with human perception, and points out that existing kinematic errors miss contact issues such as foot skating.

  1. Sunday Robotics78

    Sunday Robotics previews ACT-2 model with 99.1% zero-shot success rate in clothes folding

    Sunday Robotics previewed its robot model ACT-2, which achieved a 99.1% (±0.3%) zero-shot success rate in clothes folding across diverse unseen home environments with zero per-home adaptation, covering 9 clothing categories across 785 autonomous attempts, with an average folding quality of 4.72/5.

    Why it matters: The official disclosure provides detailed quantitative data on pretraining scale and generalization gaps, allowing readers to understand how a small amount of demonstration data transfers to unseen environments.

  1. X Square Robot/自变量机器人67

    X Square Robot releases cross-modal embodied action tokenizer X-Tokenizer, with 13.5% gain in multimodal alignment and 8.25% gain in long-horizon tasks

    X Square Robot has released X-Tokenizer, a cross-modal embodied action tokenizer built on a semantic residual quantization (SRQ) architecture and pretrained on 17 robotic arm series and 2.4 million trajectories (including 2 billion action frames). Compared with FAST, it improves multimodal alignment by 13.5% and long-horizon task performance by 8.25%, and scores 82.8 on RoboTwin 2.0; in real-robot tests it outperforms comparison methods such as FAST and 4-level RVQ.

    Why it matters: The original text presents X-Tokenizer's method design and comparative experimental data, letting readers understand the action tokenizer's multimodal alignment approach and its reusable effects.

  1. HUMANOID63

    KinetIQ Ascend uses real-world reinforcement learning to boost humanoid robot manipulation throughput and success rate

    HUMANOID has launched KinetIQ Ascend, an end-to-end visual reinforcement learning method that operates under production deployment conditions, extending its KinetIQ framework to let robots improve manipulation skills through trial and error rather than relying solely on human demonstration; the company says this is the first such method publicly demonstrated on production-line humanoid robot manipulation.

    Why it matters: The original text provides specific throughput and success-rate figures for three industrial tasks and notes a pattern of predictable performance improvement with training duration, making it useful for assessing the practical value of real-world reinforcement learning on humanoid robots.

  1. Robot.com Holdings Inc.68

    Robot.com releases R-KNav 10,000-hour sidewalk navigation dataset and launches FoMo Challenge

    Robot.com has released the R-KNav dataset and launched the FoMo Challenge, inviting the community to train visual navigation models on real data from its R-kiwi sidewalk delivery robot fleet and to evaluate them in controlled tests on physical robots.

    Why it matters: The data comes from a long-running commercial delivery fleet rather than simulation, and the KPI criteria for physical-robot evaluation have been published, making it possible to assess the reusability of real-street navigation data.

  1. Berkeley AUTOLAB68

    CaP-X: Benchmarking and Improvement Framework for Robot Manipulation Coding Agents Released

    Teams including Berkeley AUTOLAB propose CaP-X, an open framework for systematically studying Code-as-Policy agents in robot manipulation, centered on the interactive environment CaP-Gym and CaP-Bench, which evaluates 12 frontier models.

    Why it matters: The original text presents a systematic evaluation framework for Code-as-Policy agents and the finding that test-time scaling can mitigate abstraction dependence, offering reference value for research on coding-agent manipulation control.

  1. Dexterity60

    Dexterity details evaluation methods and production metrics for the Foresight packing agent

    Dexterity published an article introducing system evaluation and production measurement for the Foresight packing agent. On the planning side, it was tested on four real warehouse box-type distributions, with density of 0.65–0.77 and stability of 99.7–100%, and distribution-free order statistics were used to give 95% confidence intervals; the article also found that randomly shuffling the box sequence causes local variance to collapse by about 1.4 times, making evaluation overly optimistic.

    Why it matters: The original article fully discloses the evaluation method for the packing agent, including distribution-free confidence intervals and the evaluation bias caused by random shuffling; the method can be transferred to similar systems.

  1. Toyota Research Institute Robotics77

    Toyota Research Institute explains statistical thinking for robot policy evaluation: from rigorous A/B testing to effective visualization

    The Toyota Research Institute (TRI) published a tutorial article introducing a statistical A/B testing method based on success/failure labels for robot policy evaluation.

    Why it matters: The TRI authors systematically explain statistical pitfalls and reusable tools in robot policy A/B testing, with runnable Python examples.

  1. Physical Intelligence74

    Physical Intelligence proposes a physical intelligence layer and publishes π 0.6 deployment data from Weave and Ultra

    In a blog post, Physical Intelligence argues that robots need a reusable physical intelligence layer, and it published real-world deployment data from partners Weave and Ultra using π 0.6.

    Why it matters: The two deployers provided comparative data on autonomy rates, missed grasps, and throughput, making it easier to judge the practical gains that foundation model iterations bring in real-world settings.

  1. Physical Intelligence64

    Physical Intelligence fine-tunes π 0.6 to take on five Robot Olympics tasks, winning three golds and two silvers

    By fine-tuning its latest π 0.6 model, Physical Intelligence achieved preliminary solutions for the gold-medal tasks in 3 of the five everyday manipulation challenges in the Robot Olympics proposed by Benjie Holson, and reached silver-medal level in the remaining 2, with an average success rate of 52% and task progress of 72%.

    Why it matters: Using the same set of externally proposed everyday manipulation tasks to compare performance with and without robot pretraining lets readers judge the practical value of foundation-model pretraining for manipulation skills.

You’re all caught up