# The Pi0-Fast Libero-10 Baseline Is 25 Points Low > Source: > Published: 2026-09-17 20:03:36+00:00 ## **A duplicated transform was costing pi0-FAST 24 points on LIBERO-10** pi0-FAST is not a 60% model on LIBERO-10. It is an 85% model that has been fine-tuned incorrectly, and everyone is citing the wrong number. LIBERO actions are already deltas. The `pi0_fast_libero` config applies a `DeltaActions` transform on top of them, so the current pose gets subtracted a second time, and the model trains against a target nobody intended. `AbsoluteActions` adds the pose back at inference, so nothing crashes and the policy broadly works. It learned against a distorted distribution, and it scores like it. Someone [ran the ablation](https://github.com/Physical-Intelligence/openpi/issues/1047) and turned off the extra transform during fine-tuning. LIBERO-10 goes from 60.5% to 84.7% under mean/std normalization, and from 75.1% to 85.9% under quantile normalization. Three seeds, fifty trials per task, no model changes, and no new training code. Goal gains 5.9 to 7.5 points. Spatial and Object move 3.5 or less, so this is not a uniform lift. The 60.2% in the openpi README is the number that traveled. MemoryVLA, the Unified VLA paper, VLA-OS, and VLA-JEPA all cite it as their pi0-FAST baseline, and they are not the only ones. There is one number to check against. If your model scores below about 85% on LIBERO-10, a correctly configured pi0-FAST beats it, regardless of what your table says about the 60.2% baseline. Physical Intelligence has not responded, so this is one engineer’s ablation rather than a confirmed fix. The code is public, and the rerun takes an afternoon. ## **Research** ### **XDOF gated ABC-130k, and the LeRobot mirror kept it available** [ABC-130k](https://huggingface.co/datasets/XDOF/ABC-130k) is roughly 130,000 bimanual episodes and about 3,500 hours on the two-arm YAM rig. I pointed you at it twice in July. Ask the API for it now, and you get `gated: auto` and an empty file list. The gate covers the filenames and the download. The README returns 401. The repo was serving 292,160 downloads a month. Its `lastModified` reads 15 September, which is the only evidence of when this happened, because the history that would date it is behind the same wall. XDOF has said nothing, and I am not going to guess at why. The mirror is what saves you. `lerobot/abc_130k_v3_train` never moved. It is ungated and Apache-2.0, with 38,915 files, still serving 15,351 downloads a month, with `_val` and at least eight community copies alongside it. The June license is why. Apache-2.0 attaches to the copy when you take it, and closing the source repo does not reach back into copies already distributed. There is a second reason to care. Dyna’s DYNA-2 report draws 27 of its 39 held-out evaluation tasks from this corpus. That made it the one public route to checking Dyna’s numbers from outside, and that route now runs through the mirrors. ### **Re-running ACT’s CVAE ablation does not reproduce the 35-to-2 drop** The original ACT paper reported that removing the CVAE encoder collapsed mean success from 35% to 2%. That’s why the encoder is still carried in imitation stacks. Bo Kang [re-ran it in the original code](https://arxiv.org/abs/2609.16745) and the collapse does not appear. Training length and the checkpoint-selection rule each flip which policy wins, so the published drop is not reproducible, and its cause is still open. The sampled latent gives little reconstruction benefit at any nonzero KL weight tested, and ACT zeroes it at inference regardless. The [forensic repo](https://github.com/aida-ugent/act-cvae-forensics) is Apache-2.0 and its `make numbers` path runs with no GPU and no checkpoint download. The transferable finding is not about ACT. It is that a widely cited ablation reverses sign depending on which evaluation checkpoint you pick. ### **GSI designed a capture rig backward from the 3D math, then published its calibration error** Grounded Superintelligence released [RoboCap and the Grounded API together](https://www.gsi.company/blogs/the-robocap-for-robot-learning), a 250 g six-camera capture hat and a device-agnostic SLAM, depth, and hand-tracking stack. They claim egocentric capture is a metrology problem first. A centimeter of hand-pose error becomes a centimeter of error in the action a robot imitates. Across 150 deployed devices, factory calibration leaves a median front-pair epipolar residual of 3.8 px. Online self-calibration takes that to 0.56 px. One pixel of disparity is about a centimeter of depth at 0.65 m. Switching between the two moves 31% of in-reach pixels by more than 5%. They note Meta’s Aria drifts about 25 arcmin between its stereo cameras from someone taking the glasses off. They train the hand tracker with no motion capture. That pretraining cuts error by 17- 40% across three headsets, unlike their own rig. They also publish where they lose, including HO3D at 15.16 mm against POEMv2’s 7.72, and state that the work establishes precision rather than its effect on policy performance. If you are buying egocentric hours collected on consumer hardware, ask your vendor about the 3.8 px figure. ### **Five labs answered how little robot data you need, and none of them agree** The question showed up five times in one week, with five answers. [Reward AI](https://www.rewardai.com/blog/OM-1/) takes the extreme position. OM-1 trains only on people wearing a sensorized glove, with no teleoperation and no on-robot data anywhere. It claims a new long-horizon task from under thirty minutes of human demonstration. [HuRo](https://arxiv.org/abs/2609.10706) keeps the robot but manufactures the data, retargeting hand motion and rendering the robot where the human arms were, across 630K episodes. Scaling lifts completion from 51.5% to 80.3%, and out-of-distribution completion from 34.9% to 72.2%. [Rhoda](https://www.rhoda.ai/research/scaling-web-video-pretraining) drops the demonstrator entirely and pretrains on web video. It measures on a customer’s bearing-unpacking line, over 200 hours of real-robot evaluation. [Odyssey-3](https://odyssey.systems/introducing-odyssey-3) keeps tens of hours of teleoperation on top of a world model. GSI, above, says the source isn’t the binding constraint; the measurement is. The evidence is not symmetric. HuRo publishes a curve, an ablation, and the pipeline. Rhoda publishes two axes and no artifact. Reward AI publishes no task success rate, no baseline, and nothing to download, and its one quantified number measures a hand tracker rather than a policy. Its post also names no one. Coverage puts the team’s prior work as DexCap, HumanPlus, and ALOHA. That lineage is why you should weigh the claim at all, and it reaches you through the press rather than through them. Read Rhoda’s two axes separately, by the way. Model size moves completion from 3.7% to 84.7%; pretraining compute moves it from 57.8% to 75.3% and is flattening by the third budget. The coverage has been merging those into one number. ### **Unitree published 60 G1 datasets in four days and licensed none of them** Sixty new dataset repos appeared on the [unitreerobotics org](https://huggingface.co/datasets/unitreerobotics) between 13 and 16 September. Twenty-three whole-body teleoperation sets with Brainco hands, seventeen with Inspire hands, and twenty with the Dex1 gripper. Fifty-five hold data, 5,009 MP4 files, and 292 Parquet files in LeRobot layout, covering laundry, dishwasher loading, shelf stocking, and conveyor sorting. Five are still empty and already registering downloads, so the upload was running as I wrote this. No paper, no blog post, no announcement. Every one of the sixty has an empty license field, checked per repo. Of Unitree’s 98 earlier datasets, 64 declare Apache-2.0. This is the largest open humanoid manipulation drop of the week, and you cannot put any of it in a training set until somebody adds a license line. ### **Four bugs that do not raise, and one fix that moved the golden images** In LeRobot, `PreTrainedConfig.from_pretrained` returns a config whose `pretrained_path` is `None`, so the loader builds the processor with empty normalization stats. Raw pixels and raw positions go into the network, nothing raises, and the same rollout code scores 4.8 where a correct path scores 118 to 199. Dump the normalizer state dict after loading. An empty dict means you are running blind. The CLI is fine; this hits notebooks, serving code, and eval harnesses. In Isaac Lab, [OVRTX cameras sharing a renderer share its binding](https://github.com/isaac-sim/IsaacLab/issues/7772), so a base and a wrist camera return nearly the same stale view by default. Measured with physics frozen, consecutive frames still differ by 3.15 and 4.90 on the 0-255 scale, which should be zero. If you are generating a two-camera dataset for a VLA, check that your wrist view is actually the wrist. A [merged fix](https://github.com/isaac-sim/IsaacLab/pull/7782) to implicit actuators on ovphysx required refreshing the Shadow Hand and Kuka golden images. Policies trained on that path before 15 September were trained against the wrong dynamics. The repo’s own reference images agreed with them. And in MuJoCo, [delayed multi-input actuators read uninitialized stack](https://github.com/google-deepmind/mujoco/issues/3597): controls of 1, 2, 3 give a direct force of 6.0 and a delayed force of 501.0. ### **Bench2Dex ships trained π0.5 and GR00T N1.5 checkpoints across twelve hands** [Bench2Dex](https://arxiv.org/abs/2609.15726) is an Isaac Lab benchmark for bimanual visuo-tactile manipulation across twelve dexterous hands. A shared tactile interface lets you compare morphologies. Twenty-six tasks, synchronized visual, tactile, proprioceptive, and object-state streams. It splits perturbations into an invariance axis, where the correct action does not change, and an equivariance axis, where it must, which most robustness suites collapse into one. The useful artifact is the checkpoint repo: trained π0.5 orbax params and GR00T N1.5 safetensors per hand and task, ungated. The code is MIT. None of the three Hugging Face repos declares a license. ### **Quick hits** - `lerobot/smolvla_libero` — ships`n_action_steps=50` , equal to chunk size, so eval runs the whole chunk open-loop. SmolVLA’s own paper puts 50 actions at 51.8% against 82.8% at 10. Paired runs lose 23 points on libero_spatial. The sibling checkpoint ships 1. - **[SmolVLA to ONNX](https://arxiv.org/abs/2609.14146)** — halves p99 latency and drops LIBERO-Spatial from 70% to about 41%. A graph audit found the “INT8” and “FP16” artifacts were byte-identical FP32 graphs. The quantization never happened. - **[Neverwhere](https://huggingface.co/datasets/ziyc/neverwhere)** — 67 real scenes as Gaussian splats with aligned MuJoCo collision meshes, raw captures included, MIT. The authors argue against training only on 3DGS-generated data and show the degradation. - **[FluxVLA Engine](https://arxiv.org/abs/2609.17210)** — a same-harness LIBERO table for 13 architectures, with a downloadable checkpoint behind every cell. The spread runs from 84.7 to 98.65 on LIBERO and collapses to 57.25% at best on RoboCasa. - **[PhysBrain 1.5](https://huggingface.co/DeepCybo/PhysBrain1.5-8B)** — 2B and 8B weights, ungated, embodied pretraining supervised entirely by human interaction video. No license declared on any of the four repos. - **[Open-H-Embodiment](https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Open-H-Embodiment)** — NVIDIA spent the week auditing rather than adding. It wrote timestamp and video-kinematic sync results into the per-dataset cards. 750 hours, CC-BY-4.0, ungated. - **[FailBench](https://arxiv.org/abs/2609.03611)** — worth catching if you missed it. The best of 13 VLM failure detectors reaches 0.77 mean balanced accuracy. Models fine-tuned for the task underperform general-purpose ones. Everything drops to near chance on contact-intensive assembly. ## **Industry** ### **NOR Flash doubles again, and an AI server pulls three to five times as much** [TrendForce](https://www.trendforce.com/presscenter/news/20260914-13235.html) puts realized NOR Flash contract price growth at 100-120% across the first half of 2026, and projects another 90-110% in the second half for 256Mb and above. The low tier moves 10-20% as Chinese capacity arrives. The demand line is explicit. An AI server needs three to five times the NOR of a conventional one, and TrendForce names robots as a driver of larger firmware images. Winbond leads, and TrendForce expects it to post the strongest bit growth. MXIC is prioritizing NAND and eMMC over NOR expansion. NOR is the boot and firmware store on almost every robot controller and carrier board. It is a rounding error per unit until it doubles twice in a year, and moving firmware to a cheaper tier is a board respin rather than a purchasing decision. ### **MLPerf v6.1 puts per-accelerator inference up 5.7x in a year, with two non-NVIDIA parts available** [MLCommons](https://mlcommons.org/2026/09/mlperf-inference-v6-1-results/) published 486 results from 30 organizations. The best per-accelerator DeepSeek-R1 server figure is 5.7x better than a year ago, and the best VLM result improved 2.99x in six months. MLPerf sorts submissions by whether you can actually buy the machine. AMD Instinct MI350P, AMD Ryzen AI Max+ 395, and Intel Arc Pro B70 all submitted as available, meaning purchasable today. NVIDIA entered Rubin and Vera Rubin NVL72 as preview, which covers hardware that is not yet on sale. NVIDIA’s own [3.7x and 2.5x claims](https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference/) sit in that preview category. The same post cites the GB300 NVL72 with 99% scaling efficiency from 72 to 288 GPUs. ### **Skild put revenue on the record and broke out the mix** Deepak Pathak and Abhinav Gupta [say Skild crossed $100M in ARR](https://www.skild.ai/blogs/skild-crosses-100m-arr) ten months after first commercial deployment, across 60-plus paying customers. ARR is not recognized revenue, and the post says so. They have recognized $50M. The split is the rare disclosure. Around 90% of revenue is manipulation and 10% mobility, of which 4% is the AMR business from the Zebra acquisition. Named deployments include Sumitomo Wiring Systems and Mitsui commercial kitchens. NVIDIA’s [same-day post](https://blogs.nvidia.com/blog/skild-ai-s1-physical-ai/) adds one Skild’s own does not: a Skild, NVIDIA, and Foxconn line assembling Blackwell systems, installing a busbar and limit block, and driving sixteen screws. That post also repeats S1’s 66% success rate “at each step.” Accurate as far as it goes. It still omits that a human recovers the robot between steps. ### **Agility unveiled Digit 5, and priced nothing** [Digit 5](https://www.agilityrobotics.com/news) takes payload up 40% to 22.7 kg and moves the run-to-charge ratio from 2:1 to 10:1 on a 90-minute battery. Early access is H1 2027. The $8,500-a-month and $200,000 figures in this week’s coverage are not in Agility’s release; they come from Forbes. The “>$300M in multi-year orders” line is in the release, dated as of May. The relevant safety standards, ANSI/A3 TR R15.108 and ISO 25785-1, are still in development. Headlines calling this the first humanoid cleared to work beside people run ahead of the standard that would clear it. ### **A human-motion data vendor is priced at half a billion** Mecka pays people to record everyday tasks with body sensors and a phone, then sells the motion data for humanoid training. TechCrunch reports it is [nearing a $500M valuation](https://techcrunch.com/2026/09/11/mecka-ai-nears-500m-valuation-in-sequoia-led-deal-amid-rush-for-robot-training-data/) in a Sequoia-led round, three months after raising $60M. Terms are not final, and both sides declined comment, so treat it as reported, not announced. The valuation is not the interesting part. The interesting part is that a pure capture vendor gets priced like this inside a year. The five labs above are still arguing about whether you need the hours at all. ## **Shipped, and Still Owed** - **SolarWM’s headline 33B backbone landed.** The MiniMax-H3 build read “Coming soon” last week.[It now holds 32 safetensors](https://huggingface.co/junchaoh-cs/SolarWM-H3-33B) under a`-base` path, added on 13 September. It sits behind a gate and declares no license. The staged checkpoints above base remain invisible. - **Vention’s GRIIP SDK shipped, and it is not open source.** The CEO said Vention would open-source GRIIP and a public SDK.`griip-sdk` went to PyPI on 11 September, with`griip-core` following. Both declare`license: Proprietary` and link to no source repository. You can install it. You cannot read it. - **Dynin’s paper arrived and its artifacts did not.** The[paper](https://arxiv.org/abs/2609.13053) is real. It pretrains an omnimodal masked-diffusion VLA on 1.33M trajectories from 48 Open X-Embodiment datasets, and reports decoding up to 29.2x faster than its base implementation. The site offers the paper, model, and Code. The model link holds eight PNGs and a README. The code link holds a README and an assets folder. ## **Worth Watching** - IROS 2026 runs 27 September to 1 October in Pittsburgh. Three challenges report immediately afterward, and one hands every team roughly 7,000 hours of in-the-wild teleoperation data. - The OpenVLA fix from last issue. A [pull request](https://github.com/openvla/openvla/pull/348) landed on 16 September for the bug that made the model return the same action for a black frame and a real one. It is still open and filed by a non-maintainer. It also carries a caveat. The remote code for published checkpoints needs a matching update, so`trust_remote_code=True` loads stay broken even after this merge.