1Policy learning & VLAs

2Manipulation, dexterity & sensing

Also (all preprints): a tendon-driven five-fingered hand with dual-modality tactile per finger, PRISM GPU sampling-MPC with QP projection for bimanual contact, and on-robot juggling learned in minutes — the last a reminder that online adaptation on hardware can beat closing the sim2real gap.

3Platforms & embodiment

Thin week. Microduck (press): $399, 25 cm, 15-DOF biped from Pollen/Hugging Face, RK3566, ToF lidar, two IMUs; SDK, sim, and RL training stack open on GitHub, hardware files not. Ships before Christmas 2026. Capability shown is video only (waddling, beak pickup, fall recovery, roller skating) — demo-video evidence, staged-vs-demonstrated unknown.

SOLO (perceptive humanoid locomotion) and Poppy LQR walking on low-cost open hardware are the only platform preprints with method behind them. No humanoid company published a technical result in window.

4Open source & tooling

Friction-reducing this month, in order: CLAP’s 12 GB world model (above), which makes learned dynamics testable on hardware you own. CRESSim-Neo (preprint), GPU-batched deformable/surgical sim with PyTorch hooks. Sharpa’s tactile deformation encoder — Apache-2.0 ConvNeXtV2 weights from a sensor vendor; small, but pretrained tactile features from someone with real data are rare. Muse-Robotics-1, 121.7M-param MIT-licensed VLA, no paper, no card, unverified.

Nothing in window from LeRobot, MuJoCo/MJX, Isaac Lab, ManiSkill, RoboCasa or ROS 2 core. Genesis’ release page has 404’d for three consecutive sweeps — status unknown, not “no release.”

The eval item worth your time: statistical audit of physical-AI benchmark redundancy (preprint): 51 models × 12 benchmarks; collapsing two substitute pairs moves 22 of 51 models by three or more places, and four benchmarks retain 78.5% of the ranking utility of all twelve. Read as: most benchmark tables you see are reporting the same axis several times.

5Money & moves

EXL/iMerit folds a physical-AI annotation shop into an analytics firm; NSF put $30M over ten years into a UT Austin center on robots and people learning together. inference the data-labeling layer is consolidating into enterprise services while long-horizon HRI retreats to federal money.

6Deep read: PredVLA

arXiv:2608.26673 — preprint, Sawada & Kasahara, Sony Computer Science Laboratories.

Setup. LIBERO, all four suites, official demos, Franka in sim, 14 seeds, each suite trained separately. Held out: nothing beyond LIBERO’s own splits — no cross-embodiment, no hardware.

Method. Inherited: multiple-timescale RNN and PV-RNN’s prior/posterior free-variable formulation; frozen ResNet18/MiniLM front ends; Gaussian-mixture action head. New: (a) running it as a language-conditioned policy under the modern VLA protocol, which predictive-coding sensorimotor work had not done; (b) the strict rule that observations touch the model only through free-energy minimization over latents, which is what buys the exact open-loop ablation; (c) two lateral pathways — a low-dim visual→action bottleneck, and an efference copy of the model’s own predicted action and proprioception fed back to the visual branch, so no measurement enters forward dynamics.

The number that matters. Not the 75.35% four-suite mean — the controlled comparison. At matched parameters and identical front end, demos, action head and protocol: BC-Transformer 19.73%, BC-RNN 10.26%. Action-chunking Transformers with more parameters do worse (16.70% at chunk 8, 3.44% at chunk 16). Chunking collapses at this budget.

Where it’s soft.

• Sim only, one embodiment, one benchmark; 40.57% on LONG is not a long-horizon result.
• Per-suite training means no multi-task claim survives.
• The baselines are the authors’ own parameter-matched implementations, not published systems, with no reported hyperparameter search. A 3.4% chunk-16 Transformer is a suspiciously broken baseline — plausibly a tuning artifact of the sub-1M regime rather than evidence against chunking.
• The comparison is against small BC policies, not π0.5 or OpenVLA. The paper does not claim to beat those.
• Compute asymmetry runs in the authors’ favor, which is the honest direction.

Steal list.

1. The exact open-loop switch: if the observation pathway is separable, you can measure what closed-loop correction is actually doing on identical weights. Closing it cost 6–13 pp here. Most policies cannot answer this at all.
2. Frozen encoder + fixed PCA basis fitted on your own demos — cheap, deterministic, and it makes the learned part small enough to iterate on in minutes.
3. Efference copy as the feedback path: feed back the model’s predicted action and proprioception, not measurements, keeping forward dynamics generative.
4. Per-channel corruption sweeps. Their sharpest finding is asymmetric — 25% proprioceptive corruption cost 49.7 pp on OBJECT, 50% visual corruption under 5 pp. Worth running on anything you deploy.

Cost to reproduce. 0.68M trainable parameters, frozen front ends, public LIBERO demos: estimate single-digit GPU-hours per suite on one consumer card, no data collection, no pretraining. Test-time error regression is the only unusual cost, and 46 ms/step includes it. inference a weekend, with the risk sitting entirely in whether the baselines were fairly tuned.

7What this changes for a small build

Assume one arm, one GPU, no fleet. Three items move something this week.

PredVLA is the only paper here you could fully reproduce. 0.68M trainable parameters, frozen off-the-shelf front ends, public LIBERO demos, single-digit GPU-hours per suite. If you want a policy you can actually read end to end and instrument, this is the architecture to clone — with the caveat from §6 that the baselines may be undertuned.

CLAP makes world models testable without a cluster. Under 12 GB VRAM on a 3060 is the difference between “read the paper” and “run the thing.” If a learned dynamics model was previously out of budget, it isn’t this week.

VISTA removes a hardware dependency. If a tactile array was on your build list purely to get a contact signal into the loop, a compliant gripper plus the cameras you already have may be enough. No code released, so this is a design cue rather than something to pull, and it needs a gripper whose deformation is visible and repeatable.

Nothing this week bears on data-collection rigs or teleop cost.

8Market signal

Where headcount is visibly going: Generalist ($600M raised in total), 1X if the SoftBank deal closes, Locus (absorbing the Nexera manipulation team), and a new NSF-funded centre at UT Austin. Worth noting that this week’s manipulation-specific demand came from a logistics buyer, not a humanoid one.

Three technical clusters recur across the work that got attention. Streaming and asynchronous inference (StreamPI, FlashVLA, few-step flow sampling) appears three separate times — the field has moved from “does the policy work” to “does it run at control rate.” Contact and force as first-class policy inputs rather than safety wrappers (VISTA, TacForcing) is the second. Cross-embodiment latent-action interfaces (CLAP, camera-centric pretraining) is the third.

Table stakes, in the sense that they no longer distinguish anything: LIBERO ≥95%, action chunking, flow or diffusion action heads, a repo linked at submission. Still scarce: controlled parameter-matched ablations (almost nobody runs them, and they are why PredVLA is legible), real-robot evaluations with stated trial counts, and any evidence the authors know which benchmarks are redundant.

No credible compensation or levelling datapoints surfaced this week. The aggregator pages that rank for “robotics engineer salary 2026” are SEO content with no stated methodology; reporting nothing beats laundering them.

9Threads

First issue — board initialized.

Also seen

• GaussVLA — Mamba backbone consuming 3D Gaussian tokens. Architecture note, no eval that isolates the Gaussians.
• GaussianDream++ — compact Gaussian world-state tokens inside the VLA. Overlaps CLAP without the release.
• Zero-WAM — in-context imitation from one human video. Interesting claim, no code, unreproduced.
• R3 — RL post-training for free-form language reasoning that steers a low-level policy.
• RA-VLA — retrieval for test-time adaptation. Retrieval corpus not characterised.
• One Policy, Many Embodiments — camera-centric action geometry for cross-embodiment pretraining. Same territory as CLAP.
• LM-X — progress, event and uncertainty heads on a VLA. Interpretability claim without a decision-relevant eval.
• Memory Anchors — continual learning without full retraining. Small task suite.
• TemporalFlow-VLA — surface temporal flow as execution history. Close to StreamPI, weaker eval.
• GRAFT — cached vision-language prefix makes online VLA adaptation cheap. Worth revisiting if code lands.
• FLARE — failure detection and recovery wrapped around long-horizon VLA execution.
• WALL-SS — next-scale autoregression for longer world-model rollouts.
• 4DSynth — language or a photo into editable 4D scenes with physics.
• Figure: Index — crowdsourced smartphone data pipeline, $1B committed, nothing released.
• Jetson Orin Nano 2 — 78 TOPS, 8 GB, 2× inference at 40% less power. Ships H1 2027, so it changes nothing yet.
• Anthropic Model Hardware Standard — driver primitives over MCP for agents controlling lab arms. Research preview, nothing released.

10Open question

Is PredVLA’s 4–7× margin evidence for predictive coding, or evidence that Transformers and action chunking simply don’t function below one million parameters? The paper can’t distinguish these — every baseline is its own implementation in a regime nobody optimizes for. Settled by either a properly tuned sub-1M Transformer from an independent group, or PredVLA holding its margin at 10M–100M parameters where Transformer baselines are known-good. The second is cheap, and the authors didn’t run it.

Sources labeled inline. All arXiv items are preprints, unreviewed. Single-lab claims flagged where relevant.