Embodied AI frontier highlights, September 2026: 14 key items across eight tracks

This article is based on the archive of the Physical AI Radar. The radar has published daily since 2026-09-19 and, as of September 28, had collected 67 items (24 first-party official [O], 24 preprints [R], 19 secondhand media [M]). We picked at most two substantive items from each of the eight tracks and checked the numbers in each against the original source. The two clearest signals in September: VLA evaluation is starting to move beyond the single metric of “initial-state success rate”, and inference latency has become a metric that papers formally report.

How to read this

Evidence tags: [O] is a vendor’s or institution’s own blog or documentation; [R] is a preprint, where all numbers are self-reported by the authors and mostly not peer-reviewed; [M] is a media report, and a primary source should be found before citing it.
Track assignment follows the radar’s own results. The radar assigns tracks by keyword matching, which produces misclassifications; for example, a blog post about games launching on GeForce NOW was filed under the “training” and “simulation” tracks. Such items, as well as LLM evaluation articles unrelated to embodied AI, have been removed here.
Dates: a few items were collected by the radar in September although the paper itself was published earlier; the original date is noted in the text.

Models and training

Foundation models (VLA / WAM)

GeoAAC: adaptive action chunk length from the geometry of denoising trajectories ([R], 2026-09-17). The authors found that the geometric variation of flow matching denoising trajectories across different action prefixes correlates positively with prediction uncertainty, and use it to decide, within a single generation, how long an action chunk to execute at this step, with no extra training. On GR00T N1.5 and π0.5, simulation results improve by up to 8.7 percentage points, and the average real-robot success rate rises from 53.3% to 74.4%. A fixed chunk length is a default hyperparameter that few people touch in VLA deployment today; this work shows it can be adjusted by task phase at little extra cost.

Telexistence trials DreamZero on AWS ([O], AWS Physical AI Blog, 2026-09-09). Telexistence has deployed more than 300 robots in convenience stores. This time it ran a proof of concept with DreamZero, NVIDIA’s 14B-parameter world action model, to see whether it could beat its existing VLA. The post is blunt about the gaps: the best model in simulation reached a 45% success rate; fine-tuned on Telexistence’s real-robot data and evaluated in simulation, success dropped to 0%, which the authors trace to differences in joint distributions and state scales between the real robot and the simulation. With optimizations such as DiT caching and cutting denoising steps from 16 to 4, a B200 ran about 583 ms per step, which the authors describe as near the range needed for real-time control. It is positioned as a PoC, not a deployment, but first-hand material that spells out the shortcomings of large world models in a commercial setting is rare.

Training and self-improvement

VLA post-training for high-DoF dexterous hands ([R], 2026-09-17). The pipeline has four steps: learn a temporal hand-action codec so that an open-source VLA can output absolute dexterous-hand commands, then do SFT, DAgger and real-robot residual RL in turn. On 5 real-robot tasks, including bimanual handover, in-hand reorientation and tool use, each tested 20 times, every task reached 100% success. Its value lies in a complete, reusable post-training recipe; but each task was tested only 20 times, and the results are qualified as “within the reported post-training budget”.

Uncertainty-gated exploration noise suppresses task collapse in online RL fine-tuning ([R], 2026-09-23). During online RL fine-tuning, the overall average score can look normal while the capability on individual tasks has already been destroyed; the authors call this task collapse. The experiments run PPO on LIBERO-10 with the 450M-parameter SmolVLA: fixed noise collapsed in 2 of 3 seeds, learned noise collapsed in every seed tested, and the gated controller collapsed in none. The authors also state, however, that within this compute budget no configuration beat the behavior cloning baseline.

Data and evaluation

Data engines

MATE: a multi-agent virtual teleoperation platform ([R], 2026-09-22). Multiple operators in different locations can teleoperate whole-body humanoids simultaneously in the same physics simulation. This produced a dataset of 24.1 hours and 2,500 joint episodes covering 5 long-horizon collaborative tasks (handover, relay delivery, cooperative carrying and others). The authors report that policies transfer zero-shot to a physical humanoid without real-robot fine-tuning. This suggests that collecting multi-robot collaboration data does not necessarily require buying several real robots; but the abstract gives no real-robot success rate.

Skild AI S1: learning a new task from a single video ([O], NVIDIA Blog, 2026-09-10). S1 takes a demonstration video as a prompt and performs new long-horizon tasks of up to 10 minutes in context, without updating weights. According to the post, S1’s per-step success rate on new multi-step tasks is about 66%, versus 9% for similar systems, and one short video is worth about 380 human demonstrations, which would take 50–100 hours to collect. Note that these numbers are Skild’s self-reported results as relayed by NVIDIA’s blog, with no published evaluation protocol.

Simulation and evaluation

RoboRecover: a benchmark for recovery after execution deviations ([R], 2026-09-24). The method takes deviated states from execution trajectories, reconstructs the scene by replaying the action prefix, and then has the policy continue the original task. The benchmark has 2,000 scenarios, 1,000 each from RoboTwin and LIBERO. Results: on LIBERO, UnifoLM has an initial-state success rate of 98.83% but a recovery success rate of only 48.00%; π0.5 has a lower initial-state success rate (94.17%) but the highest recovery success rate (64.40%). In other words, a high initial-state success rate does not mean a policy can recover once something goes wrong.

Beyond Binary Success: sample-efficient and statistically rigorous policy comparison ([R], TRI and others; paper published 2026-03, collected by the radar in September). The method uses safe anytime-valid inference for sequential testing, so evaluation can stop early once there is enough evidence, and it is not limited to binary success/failure metrics. Compared with standard batch methods it cuts the number of evaluations by up to 70%; compared with sequential methods designed for binary outcomes, by up to 50%. The authors also found that fine-grained task-progress metrics distinguish between two policies faster than binary success rate does.

Deployment: edge and systems

Edge and real-time

PASSAGE: perceptive humanoid traversal in cluttered environments ([R], 2026-09-16). The authors collected 100 hours of human motion in 1,500 cluttered scenes using VR plus inertial motion capture. A flow matching planner plans at 6.25 Hz and a whole-body tracker controls at 50 Hz, all running on a Jetson AGX Orin. Across 3 seeds, increasing the training data from 6 to 100 hours raised the collision-free success rate in held-out scenes from 48.1% to 68.9%, and to 70.3% with scene augmentation; on the real robot it was tested on 50 unseen layouts. This is an example where data scaling and on-device real-time operation hold at the same time.

rMuscle: VLA inference caching that exploits cross-execution similarity ([R], 2026-09-16). At a factory workstation a robot repeats the same kind of task, so observations, action trajectories and even the model’s internal states are highly similar. rMuscle uses a Context Cache to reuse visual token outputs and an Action Cache to reuse neuron activation patterns, achieving a 1.29–1.42x speedup on an RTX 4090 and a Jetson Thor with no drop in real-robot success rate. The speedup is modest, but it needs no model changes, which makes it a good default optimization for repetitive workstations.

Systems and orchestration

Isaac ROS 5.0 ([O], NVIDIA Blog, 2026-09-22). This release supports ROS Lyrical and Ubuntu 24.04, and provides documentation aimed at AI agents plus Isaac skills (including FoundationStereo fine-tuning and pick-and-place workflows). NVIDIA says FoundationPose object tracking is up to 5.5x faster, across hardware from Jetson Orin Nano to Jetson Thor. That NVIDIA is building “letting coding agents operate the robot software stack” as a formal feature deserves more attention than the specific performance numbers.

Qualcomm announces the acquisition of PickNik ([O] + [M], 2026-09-23). PickNik’s website confirms the acquisition was announced at ROSCon Toronto. The body of Qualcomm’s announcement page is loaded dynamically and we could not read it, so the following details come from The Robot Report’s coverage ([M]): the deal has not closed and the amount was not disclosed; Qualcomm says MoveIt 1 and 2 will keep their existing licenses and community roadmap, while the commercial MoveIt Pro will be integrated with the Dragonwing platform and the Arduino VENTUNO Q. Teams that depend on MoveIt need not worry about licensing in the short term; what to watch is whether the roadmap tilts toward a single chip platform.

Safety and embodiment

Safety, permissions, and compliance

NVIDIA Halos: a layered safety architecture ([O], NVIDIA Blog, 2026-09-21). On the robotics side, the components are IGX Thor with a Functional Safety Island, Halos Core for IGX, the Holoscan Sensor Bridge, and the open-source Halos Outside-In Safety Blueprint; the industrial standards cited are IEC 61508 and ISO 13849. Be clear about the status: on the automotive side, DriveOS 6.0 has been certified by TÜV SÜD to ISO 26262 ASIL D, whereas on the robotics side IGX Thor, Halos OS and Holoscan are still in functional-safety readiness assessment with TÜV Rheinland and are not yet certified. In the archived daily picks, this is the only item in the “Safety, permissions, and compliance” track.

Embodiment and supply chain

DAWN: quadruped parkour with world-model depth denoising ([R], accepted to IROS 2026). During training, the encoder is fed noisy depth images with clean depth as the reconstruction target, and contrastive learning aligns the two latent states. As a result, filter parameters no longer need hand-tuning at deployment, and inference cost does not increase. On a Unitree Go1 it completes parkour zero-shot from raw depth: 18 cm stairs, 70 cm gaps, 45 cm steps. Earlier vision-based locomotion work rarely published its filter parameters, so this work also improves reproducibility. The other items in this track either had only media coverage with no primary source to be found (new vendor products, form-factor exploration, a 2030 humanoid field-trial plan, and so on) or were non-robotics articles misfiled here; none of them are included.

Limitations

The archive covers only about 10 days, far from “the full picture of a month”; the radar selects a limited number of items per day, so important work it missed will not appear here.
Numbers in [R] items are all self-reported by the authors, mostly from simulation benchmarks such as LIBERO and RoboTwin, and real-robot sample sizes are usually small (for example, 20 trials per task).
[O] items also include cases where a vendor relays a partner’s data (such as Skild AI S1); these are close to self-reported and are not independent evaluations.
Items with only media coverage and no primary source were not included; the one exception is Qualcomm’s acquisition of PickNik, where the parts citing media are labeled separately.