Verified Events
DateTypeCompanyHeadlineSegment
Odyssey releases Odyssey-3 foundation world model research preview
- Odyssey reports Pro scores of 66.1 on Physics-IQ Verified video-to-video and 54.7 on image-to-video using prompt enhancement and best-of-eight selection.
- The company reports first place in three of four WorldMark categories in its own evaluations.
- Robot demonstrations include recovery from missed grasps and dropped objects; physical-system adaptation uses learned action decoders or policies.
Runway introduces Praxis-1 world action model for real robots
- Noble Machines, Standard Bots and Ultra are testing the model on their hardware, including the Standard Bots RO1 arm.
- The model learns robot actions from large-scale video pretraining combined with robot interaction data.
- Public weights are planned in the coming months, following the pre-launch partner programme.
Skild AI publishes physical self-play result with humanoid transfer
- The S1-based policy was rewarded for scoring rather than for hand-designed dribbling or tackling behaviours.
- Dribbling, shielding, tackling and recovery emerged as the policy competed against successive versions of itself in NVIDIA Isaac Sim.
- Skild says it is extending the same post-training approach to everyday robot tasks beyond sport.
Light Origins releases Light-O1 embodied foundation model
- Six pretraining runs span 3.75 billion to 120 billion multimodal tokens; the largest corresponds to 100,000 hours of human action.
- Adaptation and evaluation use the Nymeria human-action data, Unitree G1 HIW-500 teleoperation data and in-house LightBot data.
- Light-O1 reports 79.3% macro success across 24 simulated RoboCasa GR-1 tasks at 50 episodes per task, and the release includes preview weights, code and a hosted playground.
LiberAI previews Liber-0 world model, topping RoboDojo benchmark
- On RoboDojo's 42-task assessment Liber-0 Preview scored an average of 30.74 with a 25.52% mean success rate, against 11.41 and 6.91% for the pi-0.5 model.
- Liber-0 Preview completed long-horizon dexterous-hand manipulation on real robots, including a brunch task spanning more than 30 dexterous subtasks; the full Liber-0 release is to follow.
Odyssey releases Odyssey-3 multi-embodiment world model
- Odyssey reports physical robot-arm skills learned from tens of hours of demonstrations and Flexion humanoid policies trained from tens of hours of teleoperation data.
- A driving policy trained on 20 hours of simulated data ran closed-loop on public roads in India and traveled about 77% as far between safety-driver interventions as policies trained on real footage.
- The drone result remains simulation-only: tens of hours of simulated flight data produced stable indoor navigation and obstacle avoidance in simulation.
Reward AI releases OM-1 cross-embodiment robot policy
- OM-1 is trained without teleoperation or on-robot experience, using images, tactile signals, inter-finger proximity and hand-pose trajectories captured through the wearable Omnibody Hand.
- The policy outputs motion direction, speed, force and grasp timing while a separate reinforcement-learned control layer compensates for robot dynamics, disturbances and inference delay.
- Reward AI says a new task, including dynamic and long-horizon manipulation, can be learned from less than 30 minutes of demonstration data.
Amap releases open ABot-Recon streaming 3D model
- ABot-Recon accepts monocular RGB video without requiring depth sensors or pre-calibrated camera parameters.
- On KITTI-02 at 504 by 280 resolution, ABot-Recon reaches 24.45 frames per second with about 6.71 GB peak memory; the stated minimum GPU is a GTX 1080 Ti.
- Inference code, evaluation scripts and pretrained weights were public at announcement; no future training-code release is treated as completed.
Perceptron releases open Isaac 0.5 embodied foundation model
- Inputs can include images, video, language, robot state and previous actions; outputs include grounded visual responses, task progress and robot actions.
- Perceptron reports training across more than 35 robot systems, 100,000 hours of robot experience, one million hours of general video and three trillion multimodal tokens.
- Weights, training and inference code, the technical report and LeRobot support were released publicly.
Generalist releases GEN-1.5 with one-shot physical-task learning
- Across 10 tasks, Generalist reports 59% average success from one 3–12 second in-context demonstration and 83% after 10 gradient steps on five minutes of task data.
- The multimodal model maintains 30 seconds of video context alongside sensor, language and proprioceptive inputs, and produces action trajectories at 100 Hz.
- The release also shows prompts transferring from simulation to a physical robot and, in some cases, from human-hand demonstrations to immediate robot execution.
Skild AI publishes S1 in-context robot learning model
- S1 applies in-context learning to seen and unseen manipulation tasks, including plant potting, pancake preparation, pour-over coffee and kit assembly.
- Skild reports unseen tasks lasting up to 10 minutes and a sevenfold improvement over language-only prompting on its evaluation set.
- The official release demonstrates transfer from a human video example to robot execution, including an 11-minute demonstration-to-execution plant-potting workflow.
Odyssey releases CaliBench physical-calibration benchmark
- The benchmark evaluates physical calibration rather than only visual plausibility.
- It targets model use in robotics, science, healthcare, education and gaming.
- The release is a research benchmark and does not establish real-world robot deployment.
Generalist reports up to 10–20× gains in GEN-1 actuator adaptation
- The update targets low-level adaptation to new actuators and robot embodiments rather than a new GEN-1 model release.
- Generalist reports gains of up to 10–20 times on internal benchmarks, without publishing the benchmark protocol or absolute scores.
- The company showed the improvement on high-precision disassembly of parts from a NIST board.
Light Origins publishes Light-Loco-Parkour for whole-body humanoid locomotion
- The policy uses only onboard depth sensing and a velocity command to select walking, balancing, climbing, stepping down or vaulting without a runtime skill label, state machine or motion graph.
- A Real2Sim2Real pipeline expands sparse human-motion seeds across terrain variations, unifies multiple experts and distils privileged geometry into a recurrent depth policy.
- The same policy transferred without real-world fine-tuning to indoor and outdoor tests on Lightbot 0, including unseen obstacle variations.
Being-H0.8: A Latent Tactile World-Action Model at Scale
- Being-H0.8 extends latent world-action modelling with tactile-aware interaction using contact and proximity supervision derived from large-scale human video.
- TopoHand maps human hands, dexterous robot hands and parallel grippers into a shared state-action representation for cross-embodiment learning.
- BeingBeyond shows ten physical contact-rich manipulation demonstrations on two bimanual robot configurations; the accompanying paper remains forthcoming.
Alibaba AMAP releases ABot-World-0 interactive world model
- The model produces 720p interactive video at up to 16 FPS with about 1.2 seconds of latency and roughly 19 GiB peak VRAM on an RTX 5090.
- ABot-World-0 accepts raw keyboard actions for roaming and character interaction rather than relying on a fixed scripted trajectory.
- The release includes a technical report, project page, public code and downloadable model weights.
Flexion presents a Niantic Spatial and NVIDIA sim-to-real pipeline for humanoids
- Niantic Spatial reconstructed the office as a metric digital twin from camera capture.
- NVIDIA Isaac Sim and Isaac Lab provided the simulation and training environment.
- The RGB navigation policy transferred zero-shot to a physical humanoid in the real office.
Flexion releases Reflect v1.0 for long-horizon humanoid autonomy
- Reflect v1.0 uses a custom VLM mission controller to reason over instructions, scene state and mission progress.
- The stack combines semantic mapping, VLA/RL motion skills, navigation, manipulation, whole-body control and runtime safety checks.
- Flexion demonstrated autonomous multi-floor parcel retrieval, elevator/stair navigation, box opening and snack placement without a human operator.
Alibaba releases Qwen-Robot embodied-AI model suite
- The suite includes Qwen-RobotNav, Qwen-RobotManip and Qwen-RobotWorld components.
- The release targets robot navigation, manipulation policy development and simulation workflows.
Kairos World Model
- Open repository and Hugging Face organization provide Kairos world-model resources for embodied agents.
- Positions the model around multimodal understanding, generation, action prediction, and long-horizon robot reasoning.
- June 15 coverage highlights benchmark performance across RoboTwin, LIBERO-Plus, WorldModelBench Robot, and DreamGen.
Physical Intelligence publishes pi0.7 steerable robotic foundation model
- Physical Intelligence published pi0.7 as a steerable robotic foundation model with stronger generalization and task recombination abilities.
- The official release compares pi0.7 against specialist policies on laundry folding, espresso making, and box folding while highlighting cross-platform generalization.
- The release is PI's most recent first-party generalist model release and belongs in the company research tab rather than only the generic timeline.
Ludo Robotics releases Prompt-to-Policy reinforcement-learning system
- The system writes reward functions, trains policies with PPO and evaluates rollouts using code and vision-language judges.
- Its release includes a web dashboard and support for MuJoCo environments, with optional IsaacLab integration.
- The official technical report demonstrated 64 learned behaviors and published the full training and evaluation repository under an MIT license.
Spirit AI releases Spirit-v1.5 fine-tuning code
- Spirit AI updated the Spirit-v1.5 repository in April 2026 with fine-tuning code, separating the training release from the January initial model and benchmark release.
- The repository documents training requirements, dataset loading, distributed training utilities, and a fine-tuning launcher for adapting Spirit-v1.5 to robotic manipulation datasets.
- The release matters because it makes Spirit-v1.5 more reproducible as a developer resource for embodied-AI research rather than only a checkpoint and benchmark artifact.
Physical Intelligence publishes RLT research on efficient online RL for precise robot manipulation
- Physical Intelligence published RLT, an online reinforcement-learning method for precise robot manipulation.
- The work focuses on improving task performance efficiently through real-world policy updates.
Physical Intelligence publishes Multi-Scale Embodied Memory for long-horizon VLA tasks
- Physical Intelligence published Multi-Scale Embodied Memory as a way to give VLA policies both short-term and long-term task memory.
- The official page frames MEM as enabling complex tasks lasting more than ten minutes, with high-level subtasks selected at lower frequency and actions selected at high frequency.
- Memory is a core missing capability for robots expected to complete long household and workplace workflows.
ACE Robotics publishes ACE-Brain-0 foundation model research
- ACE Robotics published ACE-Brain-0, a spatial-intelligence foundation model for robotics, autonomous driving and UAV embodiments.
- The model used spatial reasoning as a shared scaffold across different physical systems.
- The research extended ACE Robotics' robot-brain work into multimodal foundation-model development for embodied systems.
Robbyant publishes LingBot-VLA pragmatic foundation model with open weights and benchmark data
- Robbyant published LingBot-VLA, a VLA foundation model trained on about 20,000 hours of real-world data from nine dual-arm robot configurations.
- The paper reports evaluations across three robotic platforms and releases code, base model weights, and GM-100 benchmark data for reproducible comparison.
- LingBot-VLA is a first-party Robbyant research release and one of the stronger open embodied-AI benchmark/data contributions in the humanoid-adjacent stack.
Spirit AI open-sources Spirit-v1.5 robotic foundation model
- Spirit AI released Spirit-v1.5 as an official VLA robotic foundation model implementation with model architecture, inference wrapper, RoboChallenge integration and model checkpoints.
- The repository says Spirit-v1.5 ranked first on the RoboChallenge Table30 benchmark as of January 11, 2026, and links the Hugging Face model release.
- The release matters because it gives Spirit AI a first-party model artifact behind its embodied-intelligence stack and provides a public benchmark path for the Moz1 product story.
Physical Intelligence studies emergence of human-to-robot transfer in VLA models
- Physical Intelligence published research on how transfer from human videos to robotic tasks emerges as VLA models scale.
- The official page studies when human-video pretraining begins to help robotic task performance rather than staying disconnected from robot action data.
- Human-to-robot transfer is central to scaling robot datasets beyond expensive teleoperation alone.
Moz1 arm paper presents analytical inverse kinematics for Spirit AI's 7-DOF robot arm
- The Moz1 inverse-kinematics paper presents a closed-form analytical solution for the seven-DOF Moz1 robot arm with wrist offsets.
- The author reports full self-motion support, singularity handling, and all 16 solutions per pose for the arm configuration.
Physical Intelligence publishes pi-star-0.6 VLA that learns from experience
- Physical Intelligence published pi-star-0.6 as a VLA trained with Recap-style reinforcement learning from autonomous robot experience.
- The release reports improvements on real-world application tasks such as box building, kitchen cleaning, and making coffee.
- The release moves the PI research timeline from imitation-style policy learning toward policies that improve through deployed experience.
Google DeepMind launches Gemini Robotics 1.5 for agentic physical tasks
- Google DeepMind introduced Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 for robots that can perceive, plan, use tools and act across multi-step tasks.
- Gemini Robotics 1.5 learns across embodiments and turns visual information plus instructions into robot motor commands.
- The model family is positioned for physical agents and humanoid partners such as Apptronik Apollo.
Google DeepMind introduces Gemini Robotics On-Device for local robot control
- Google DeepMind introduced Gemini Robotics On-Device as an efficient VLA model optimized to run locally on robotic devices.
- The model supports general-purpose dexterity, fast task adaptation and SDK-based evaluation in MuJoCo with small demonstration sets.
- Local inference is central for practical robot autonomy in latency-sensitive or disconnected environments.
Physical Intelligence publishes real-time action chunking for high-latency VLA control
- Physical Intelligence published Real-Time Action Chunking as a method for keeping VLA-controlled robots precise while large models think over future action chunks.
- The official page frames the approach as an inpainting problem that keeps new action chunks consistent with actions already being executed.
- Inference latency is a practical blocker for deploying large robot foundation models on physical robots.
Physical Intelligence publishes VLA training method for faster training and stronger generalization
- Physical Intelligence published a VLA training method focused on faster training, fast inference, and improved semantic generalization.
- The official page describes training with pi0-FAST action tokens, general web-data representation learning, and continuous actions through the action expert.
- The release fills an important research step between pi0.5 and later PI model releases.
Physical Intelligence publishes pi0.5 VLA for open-world generalization
- Physical Intelligence published pi0.5 as a VLA model focused on open-world generalization across messy homes and unstructured tasks.
- The official writeup emphasizes co-training on heterogeneous data so the model can transfer knowledge across robots, instructions, and semantic task contexts.
- The release marks a clear step from lab dexterity toward general-purpose household and workplace robot behavior.
FieldAI Research Institute and NVIDIA build embodied Q&A dataset in Omniverse
- FieldAI Research Institute announced research with NVIDIA on robotic question answering in dynamic simulated environments.
- The work uses NVIDIA Omniverse tooling to create a long-term embodied Q&A dataset for studying robot reasoning in changing scenes.
- It is a first-party FieldAI research release that belongs with FieldAI company research coverage.
Physical Intelligence open-sources pi0 weights and openpi repository
- Physical Intelligence released code and model weights for pi0 through its experimental openpi repository.
- The release includes base pretrained checkpoints, fine-tuned checkpoints for platforms such as ALOHA and DROID, and example code for real and simulated robots.
- The release made a general-purpose robot foundation model available for researchers and developers to adapt to their own platforms.
Physical Intelligence releases FAST robot action tokenizer for efficient VLA training
- Physical Intelligence published FAST, an efficient robot action tokenizer designed to connect high-frequency robot action chunks to autoregressive transformer training.
- FAST was trained on one million real robot action sequences and supports faster pi0-FAST policy training while retaining dexterous-control capability.
- Tokenizer and dataset infrastructure are part of the research substrate behind scalable robotic foundation models.
Physical Intelligence introduces pi0 general-purpose robot foundation model
- Physical Intelligence introduced pi0 as its first generalist robot policy, combining a vision-language model backbone with action generation for dexterous robot control.
- The official release describes training across multiple robot forms and tasks including laundry folding, table cleaning, grocery bagging, and box assembly.
- Pi0 is one of the core first-party foundation-model releases in the modern robotics stack.
Google DeepMind introduces RT-2 vision-language-action model
- RT-2 was presented as a robotics model trained on both web-scale vision-language data and robot action data.
- The release described improved generalization on emergent robotic skills compared with RT-1 and visual-language baselines.
- The event is sourced to Google DeepMind's official RT-2 research blog.
OpenAI trains robotic hand to solve a Rubik's Cube with Automatic Domain Randomization
- OpenAI trained neural networks to solve a Rubik's Cube with a human-like robotic hand using reinforcement learning in simulation.
- The project introduced Automatic Domain Randomization to create progressively harder simulated environments for sim-to-real transfer.
- The system solved simpler scrambles 60 percent of the time and maximal-difficulty scrambles 20 percent of the time, making it a durable dexterous-manipulation milestone.
OpenAI trains Dactyl robotic hand system for dexterous manipulation
- OpenAI introduced Dactyl, a system that trains a Shadow Dexterous Hand entirely in simulation and transfers the policy to real hardware.
- The work used the same general-purpose reinforcement learning system behind OpenAI Five to manipulate objects into target orientations.
- The event remains a landmark robotics result because it showed sim-trained dexterous manipulation could adapt to real-world physics.