Radar — Queue / Watchlist¶
⏳ Review needed
This page's updated (2026-09) exceeds the volatility 'high' threshold (1 months) by 1 month(s). The Korean source needs review; refresh updated after syncing. (update rules)
Last updated: 2026-09 · owner: Youngjin · volatility: high ← back to index
L0 TL;DR: Candidates needing further source, reproduction, or site checks. Promotion requires all inclusion criteria and owner review of intended use. Scans cannot approve promotion.
⚠️ Do not present items here as "mature capabilities" in customer proposals. A flashy demo often masks how deployable something actually is.
Review scope: the page edit date does not revalidate every technical item. See Evidence for core corrections, check dates, and reproduction/human review status; legacy item dates still apply.
🔬 Models / algorithms (awaiting validation)¶
| Item | Label | At a glance | Promotion condition |
|---|---|---|---|
| Physical Intelligence π0.7 | 🔵 Research | ✨ Notable: PI, a leading VLA lab (π0/π0.5), officially demoed compositional generalization (completing untrained tasks, e.g. cooking in an air fryer) in 2026-04 — weight release or GA could move the industry baseline again ⏳ On hold: Official PI announcement (2026-04-16, multiple outlets quote the PI blog and Sergey Levine) [4] — a research demo only, weights not released (openpi covers only π0/π0.5), and PI itself admits limits on complex multi-step instructions. 2026-10-24: corrected the earlier "rumored next flagship" wording to "confirmed official research demo" (this session could not reach physicalintelligence.company directly due to a DNS block — cross-checked via multiple independent outlets) 2026-11-04: checked the primary page (pi.website/blog/pi07, "Published April 16, 2026") and the openpi README (released up to π0.5) directly — the official announcement is [1]; performance claims remain unverified |
Weight release or GA + independent performance validation |
| World-action models (DreamZero → GR00T N2) | 🟡 Preview | ✨ Notable: The "world models that also generate actions" track touted as the post-VLA generation — a directional indicator for NVIDIA's roadmap ⏳ On hold: GR00T N2 "expected end of year," DreamZero is research |
GA + real deployment case |
| Google DeepMind Genie 3 (world model1 for robot learning) | 🟡 Preview | ✨ Notable: An attempt to use a frontier-grade world model as a data source for robot policy learning — would bypass the real-data bottleneck if it works ⏳ On hold: The world model itself is preview; applying it to robot learning is research |
Validated case of robot policy learning |
| VLM-based SysID2 (Vid2Sid, Swim2Real) | 🔵 Research | ✨ Notable: Estimates physical parameters from video alone to automate simulator calibration — could eliminate manual sim-to-real calibration ⏳ On hold: 2026 preprints, single lab |
peer-review + reproduction |
| VIRAL / VideoMimic / Real2Render2Real (visual sim-to-real3 at scale) | 🔵 Research | ✨ Notable: Visual sim-to-real that reconstructs simulation environments and demonstrations from ordinary video — a candidate to change the data-collection cost structure ⏳ On hold: CVPR/CoRL research, not production |
Evidence of production deployment |
| Robbyant LingBot-VLA / UnifoLM-VLA-0 | 🔵 Research | ✨ Notable: A new wave of open VLA families from China — watched for the open-weights competitive landscape ⏳ On hold: Secondary sources, no validation |
Primary confirmation + AWS mapping |
🖥️ Simulation / tools (awaiting maturity)¶
| Item | Label | At a glance | Promotion condition |
|---|---|---|---|
| Genesis physics engine4 | ⚪ Hype | ✨ Notable: Hyped as an "ultra-fast general-purpose physics engine" — if true, it changes the cost structure of GPU simulation ⏳ On hold: "430,000×" refuted [1], slow on contact-rich manipulation |
Independent benchmark + production adoption |
| MuJoCo Warp | 🟡 Alpha | ✨ Notable: Combines MuJoCo accuracy with GPU parallelism — an alternative candidate to the Isaac monoculture ⏳ On hold: PyPI classifier "3-Alpha" [1], not production |
Beta/GA transition |
| NVIDIA Newton physics engine | 🟡 Preview | ✨ Notable: Next-generation open-source physics engine co-developed with Google DeepMind and Disney Research — the likely next standard of the Isaac ecosystem ⏳ On hold: Experimental backend in Isaac Sim 6.0 |
GA + official in Isaac Lab 3.0 |
| Isaac Sim 6.0 | 🟡 Preview | ✨ Notable: A next-generation overhaul including Newton integration — a directional indicator for migrating the current 5.x stack ⏳ On hold: "Early Developer Release," API in flux (latest GA is 5.1) |
6.x GA declaration |
| Cosmos 3 as sim-to-real training source | 🟢 GA(model)/🔵(in practice) | ✨ Notable: The track of training deployable policies from world-model-generated data — would reshape the SDG pipeline landscape if it works ⏳ On hold: Model is GA, but "training deployable policies from world-model data" is early-adopter only. ⚠️ not hosted on AWS |
Stronger AWS mapping + training validation |
🤖 Hardware / deployment (roadmap / demo)¶
| Item | Label | At a glance | Promotion condition |
|---|---|---|---|
| Tesla Optimus V3 | ⚪ Hype | ✨ Notable: The highest-profile humanoid mass-production plan — the most frequently asked-about item by customers ⏳ On hold: Musk claims only, production not started |
Validated deployment |
| Hyundai·BD electric Atlas | ⚪ Roadmap | ✨ Notable: Hyundai Motor Group's mass-production roadmap (30k/yr from 2028) — the most direct humanoid track for customer conversations in Korea ⏳ On hold: Electric Atlas product version unveiled (2026-07, BD official [3]). Deployment 25k+ units & 30k/yr capacity both start 2028; ~0 in live production today. 2026 is a small pilot only (Hyundai RMAC + Google DeepMind). ⚠️ "Gen 5" is a misnomer |
Verified real-operation shipments begin |
| Apptronik Apollo 2 + Robot Park | 🟡 Pilot | ✨ Notable: Live operational pilots at Mercedes and GXO plus a Google DeepMind data partnership — a front-line indicator of humanoid commercialization ⏳ On hold: Operational pilots at Mercedes-Benz & GXO [3] + Google DeepMind Gemini Robotics data partnership (90k sq ft). Autonomy/commercial scale unverified. AWS mapping is generic (data→S3/SageMaker); the partnership itself is Google [4] |
Commercial deployment scale + validated autonomy |
| 1X Neo autonomy | 🟡 Preview | ✨ Notable: Among the first home humanoids actually on sale ($20k) — a test bed for the mixed-teleoperation operating model ⏳ On hold: Mixed autonomy + VR teleoperation (Expert Mode) — acknowledged by the CEO directly (Engadget [3]). The "60–70% autonomy" figure has no primary source [4] |
Validation of true autonomy |
| Figure 03 "8-hour autonomous shift" | ⚪ Hype | ✨ Notable: An autonomy claim on top of a validated BMW pilot record — would reset the bar for industrial humanoid autonomy if true ⏳ On hold: CEO tweet, no independent validation (Figure 02@BMW is a validated pilot) |
Third-party autonomy audit |
| Cosmos 3 adoption (Doosan/LG/Samsung) | 🟢 GA(announced) | ✨ Notable: Adoption announcements by three major Korean conglomerates — a reference that comes up immediately in customer conversations in Korea ⏳ On hold: Adoption is "announced," not production-validated |
Public production case |
🔗 Agents / connectivity (early)¶
| Item | Label | At a glance | Promotion condition |
|---|---|---|---|
| MCP5 for robotics (ros-mcp-server, etc.) | 🔵 Research | ✨ Notable: A surge of experiments (50+ servers) wiring the agent-standard protocol to robot skills — an AgentCore-integration angle ⏳ On hold: 50+ servers exist but open-source/demo, none in production (safety, latency, determinism unvalidated) |
Production-hardening case |
| ROS 26 + LLM agents7 (NASA JPL ROSA, RAI) | 🔵 Research | ✨ Notable: Real-organization validation cases such as NASA JPL's ROSA — the most realistic entry path from natural language to robot operations ⏳ On hold: ROSA (JPL) is the strongest real case but mock-ops. Field deployment limited |
Field production deployment |
| Agent physical-safety standards (RoboGuard, etc.) | 🔵 Research | ✨ Notable: A standards vacuum for LLM semantic-level risk — could surface as a regulatory/procurement requirement ⏳ On hold: ISO covers physical only; no standard for LLM semantic risk |
Progress on standardization |
| AgentCore Payments / Agent Registry (Seoul) | 🟡 Preview/unavailable | ✨ Notable: The AWS-native track for robot-agent commerce and registry infrastructure — immediately proposable once it opens in the Seoul region ⏳ On hold: Not available in Seoul region — Agent Registry is in Tokyo ✅, but Payments is not in Tokyo either (Sydney only in APAC) [1] |
Seoul region expansion |
🆕 Latest scan intake (2026-11-04 · links and primary sources checked 2026-09-27)¶
| Item | Label | Gist | Promotion condition |
|---|---|---|---|
| NEURA Robotics 4NE1 / Neuraverse × AWS (German full-stack robotics, strategic collaboration with AWS) | ⚪ Roadmap | ✨ Why it matters: AWS serves as the primary cloud provider hosting the Neuraverse platform, NEURA Gym training pipelines integrate with SageMaker, and NEURA expands go-to-market through the AWS Partner Network (APN) — an AWS partnership with named services, the European humanoid track on the Radar ⏳ Why waiting: official AWS–NEURA announcement (2026-04-21, press.aboutamazon.com) [4] — Amazon fulfillment-center deployment is worded as "explore opportunities", not a live deployment. Momentum from the Series C of up to $1.4B (2026-06-10; Amazon, NVIDIA, Tether and others; company wording "largest ever for a full-stack robotics company") and the IFA 2026 Berlin keynote (2026-09-05); no third-party validation. 🔗 Links 200 · primary sources checked 2026-09-27. AWS angle: official collaboration (SageMaker and APN named) |
Live deployment cases (e.g., Amazon fulfillment centers) + independent performance validation |
| RLWRLD RLDX-1 (8.1B-parameter dexterity8 foundation model trained on AWS) | 🔵 Research | ✨ Why it matters: an open robotics foundation model from a KAIST-founded Seoul startup, featured directly by the AWS Physical AI Blog (2026-06-22) — trained on EC2 p5e/p5en (H200), ParallelCluster and FSx for Lustre, specialized in five-finger precise manipulation, code and weights released on GitHub (CC BY 4.0) — with NEURA, an official AWS collaboration case on the Radar and a Korea-originated track. A second AWS blog post followed on 2026-10-28 (partnership status and Marketplace listing — see News) ⏳ Why waiting: arXiv technical report (2026-05-05) + AWS blog [4] — self-measured results on 6 simulation suites and 3 real robot platforms (GR-1 Tabletop 58.7 vs GR00T N1.6 47.6 = +11.1 pts, Table 1(b); ALLEX real-robot 86.8% vs about 40% for π0.5 and GR00T N1.6; SIMPLER 81.5% as quoted by the AWS blog), no independent reproduction or peer review — self-reported figures, do not quote to customers. The AWS relationship is training infrastructure plus Generative AI Accelerator participation (2025-10); the deployment target is LOTTE Hotels & Resorts by 2030, subject to safety validation. 🔗 Links 200 · primary sources checked 2026-09-27. AWS angle: AWS blog case (EC2, ParallelCluster, FSx) · 🇰🇷 Korea touchpoint |
Independent benchmark reproduction + published deployment case |
| UniWALM (a unified World-Action-Language model that co-generates video, action, and language in a single diffusion backbone and re-consumes the generated language as next-step memory) | 🔵 Research | ✨ Why it matters: extends a World-action model (video + action) by folding language generation into the same backbone, closing a "re-feed the generated language into the next step" memory loop with no separate planner or memory module — most authors are Amazon-affiliated, adding an Amazon-affiliated research case to the Radar's data/policy-unification axis ⏳ Why waiting: arXiv preprint (2026-10-20) [4] — the reported gains on memory-dependent tasks (part of RMBench, and the real-robot Find-the-Item task, where success rose from 11.1% to 61.1%) over the same-backbone baseline are entirely the authors' own measurements, limited to a single group and a handful of tasks, with no independent reproduction or peer review — self-reported figures, do not quote to customers. 🔗 The link could not be curl-200 verified this run due to a blanket egress block (even DNS for arxiv.org failed) — cross-checked via the alphaXiv original (ID 2610.28626) instead. AWS angle: Amazon's own research, no AWS service tie-in |
Independent reproduction + validation across more tasks and platforms |
| LG×NVIDIA humanoid MOU (LG's bipedal humanoid built on NVIDIA Isaac GR00T, Jetson Thor and Halos for Robotics, targeting a Q1 2027 unveiling) | ⚪ Roadmap | ✨ Why it matters: LG announced a group-wide effort (LG Electronics actuators, LG Innotek sensors, LG Energy Solution batteries) to build its own bipedal humanoid on NVIDIA's Isaac GR00T, Jetson Thor and Halos safety stack, via an MOU signed 2026-08-13 — a distinct track from the Radar's existing "Cosmos 3 adoption" item (Doosan/LG/Samsung, a software-adoption announcement): here LG commits to building the humanoid hardware itself, a track likely to come up in Korean customer conversations ⏳ Why waiting: only an LG–NVIDIA MOU and development plan [4] — the public unveiling is targeted for Q1 2027 (CES mentioned), with zero units in operation today. A companion in-house robot foundation model (RFM) is also roadmap-stage. 🔗 Link 200 · checked locally 2026-10-24 (PR Newswire original title matches) — content rests on the cross-check across multiple independent outlets (Korea JoongAng Daily, eWeek, Digital Trends, etc.) made at intake. AWS angle: none (NVIDIA–LG bilateral collaboration) · 🇰🇷 Korea touchpoint |
Public unveiling (Q1 2027) + independent performance validation |
| AWS Physical AI Toolchain (AWS's official open-source Physical AI reference architectures, IaC and deployment automation integrating NVIDIA Isaac/GR00T/Cosmos) | 🟢 GA (open source) | ✨ Why it matters: AWS published robot- and task-agnostic Physical AI reference architectures (Terraform/CDK) and deployment automation under the official aws-samples org — wiring EC2 (G6e/P5), SageMaker, EKS, AWS Batch (EFA), S3 and FSx for Lustre to NVIDIA Isaac Sim, Isaac Lab, GR00T (N1.6), Cosmos and OSMO — the most direct official AWS case tying together several Radar axes (World-action models, Isaac Sim 6.0, Cosmos 3, GR00T licensing) on AWS infrastructure ⏳ Why waiting: official AWS blog post (est. 2026-10-07/08, aws.amazon.com/blogs/physical-ai) + public GitHub release (Apache 2.0) [4] — an "aws-samples" repo with no formal support commitment; the README's component table marks Isaac Lab Arena Evaluation as Preview and Jetson edge deployment as Planned, and no live customer production-adoption case has been confirmed. 🔗 Link 200 · checked locally 2026-10-24 — official GitHub repo plus the official AWS blog post reached directly (not verifiable in the intake run due to egress limits, issue #20). AWS angle: AWS's own official open source (top tier) |
Published live customer production-adoption cases + primary-source verification of the official AWS blog |
| Real-Time EXPO-FT (Stanford; a latency-aware real-time policy-correction framework that compensates for a large VLA's asynchronous inference delay using a lightweight correction policy and a chunk-level RL critic) | 🔵 Research | ✨ Why it matters: even with up to 167ms of inference delay, a π0.5-based VLA that asynchronously pre-generates action chunks for the delay window and lets a lightweight policy correct them against the latest observation raises real-robot success on dynamic manipulation (ball catching, soccer kicking, etc.) from 12.5/30 to 29/30 with under 10 minutes of online robot data — adds a new "inference latency / real-time" axis to the Radar's 🔬 VLA track; the highest-engagement robot-learning paper in today's scan (273 votes / 18,981 views on alphaXiv, far above any existing intake item) ⏳ Why waiting: arXiv preprint (2026-09-16) [4] — a single Stanford lab's own measurements, limited to 4 real-robot tasks on a single robot platform, dependent on task-specific success detectors for sparse binary rewards, with human-assisted environment resets, and no independent reproduction or peer review — self-reported figures, do not quote to customers. 🔗 The link could not be curl-200 verified this run due to a blanket egress block (DNS for arxiv.org failed) — cross-checked via the alphaXiv original (ID 2609.18207) instead. AWS angle: none (a single Stanford research effort, no AWS service tie-in) |
Independent reproduction + validation on the full VLA pipeline (currently verified mainly via a state-based flow-policy simulation) |
| SimVLA (a joint Yonsei University/Seoul National University project that pre- and post-trains a VLA entirely on simulation data, with no teleoperation, for zero-shot sim-to-real transfer in mobile manipulation) | 🔵 Research | ✨ Why it matters: a Korean-university collaboration (Yonsei, SNU) trains a VLA purely on simulation-only supervision — action trajectories, VQA-style visual-language annotations, and policy rollouts — and transfers it to real home environments with no additional training, a Korea-originated case testing whether mobile-manipulation sim-to-real is possible without any real-world data collection, adding a Korean-academia track to the Radar's 🔬 visual sim-to-real axis ⏳ Why waiting: arXiv preprint (2026-10-08) [4] — the claim that it outperforms a π0.5 baseline fine-tuned on 50 real demonstrations, including in real-home transfer, is entirely the authors' own measurement, limited to a single lab and a handful of tasks (restocking, pouring, cleaning), with no independent reproduction or peer review — self-reported figures, do not quote to customers. 🔗 Link 200 · checked locally 2026-10-24 (arXiv title matches) — content rests on the alphaXiv cross-check made at intake (the intake run's egress block prevented curl). AWS angle: none · 🇰🇷 Korea touchpoint (Yonsei, SNU) |
Independent reproduction + validation beyond the home setting |
| Mulligan (a performance-guided robot data-collection strategy that picks the next teleoperation episodes' starting states around where the policy still fails, instead of uniform sampling) | 🔵 Research | ✨ Why it matters: shifts teleoperation data collection from "allocate budget proportional to the reset distribution" to "prioritize states where the policy still fails," raising real-robot success rate at the same collection budget — adds a "what and how to collect next" strategy axis to the Radar's data-pipeline track (FIBER, UniWALM, etc.); the highest-engagement robot-learning paper in today's scan (60 votes / 394 views on alphaXiv) ⏳ Why waiting: arXiv preprint (2026-10-05) [4] — the real-robot success-rate gains (3 Franka Panda tasks; the final-round INSERT MARKER +34 pts is statistically significant, but the final-round THREAD NUT and ROUTE CABLE comparisons are not) are entirely Stanford's own measurements, limited to a single robot platform (DROID) and 3 tasks, with no independent reproduction or peer review — self-reported figures, do not quote to customers. 🔗 The link could not be curl-200 verified this run due to a blanket egress block (even DNS for arxiv.org failed) — cross-checked via the alphaXiv original (ID 2610.05882) instead. AWS angle: none (a single Stanford research effort, no AWS service tie-in) |
Independent reproduction + validation across more platforms and tasks |
| ThinkWorld (a joint NVIDIA/KAIST Chain-of-Dynamics framework that supervises even unlabeled human video via reinforcement learning on whether the predicted future world state matches what was observed, improving VLA human-to-robot transfer and generalization to unseen tasks) | 🔵 Research | ✨ Why it matters: complements the limited diversity of robot-only demonstration data with human video (EgoDex, EgoVerse), supervising even action-unlabeled video directly via future-state prediction accuracy, and demonstrates human-to-robot transfer and generalization to unseen tasks in both simulation and on real robots — the highest-engagement robot-learning paper in today's scan (53 votes / 595 views on alphaXiv) ⏳ Why waiting: arXiv preprint (2026-10-27) [4] — a joint NVIDIA/KAIST paper; the reported generalization results are entirely the authors' own measurements, limited to simulation plus a handful of real-robot platforms, with no independent reproduction or peer review — self-reported figures, do not quote to customers. 🔗 The link could not be curl-200 verified this run due to a blanket egress block (DNS/connection failures for both arxiv.org and alphaxiv.org) — cross-checked via the alphaXiv original (ID 2610.37666) instead. AWS angle: none (NVIDIA/KAIST joint research, no AWS service tie-in) |
Independent reproduction + validation across more platforms and tasks |
| FLORE (Amazon FAR) (a force-reactive loco-manipulation framework that uses UMI9-style robot-free demonstrations to record contact force alongside motion, training a humanoid to control force and whole-body motion together) | 🔵 Research | ✨ Why it matters: where whole-body control has typically tracked motion (position) alone, this records, predicts and tracks contact force from the demonstration onward, enabling contact-rich tasks like rotating a box or wiping a board — adds a new "force control" case to the Radar's Amazon FAR research axis after UniWALM ⏳ Why waiting: arXiv preprint (2026-10-26) [4] — the success rate on 3 real-robot tasks on a Unitree G1 (pushing a gantry, rotating a box, wiping a board; 10 rollouts per task) of 80% vs. 26.7% for a position-only baseline is entirely Amazon FAR's own measurement, limited to a single platform and 3 tasks, with no independent reproduction or peer review — self-reported figures, do not quote to customers. 🔗 The link could not be curl-200 verified this run due to a blanket egress block (DNS/connection failures for arxiv.org; alphaxiv.org also unreachable) — cross-checked instead by reading the full alphaXiv original (ID 2610.35862, authors, method and figures). AWS angle: Amazon's own research (FAR), no AWS service tie-in |
Independent reproduction + validation across more platforms and tasks |
Retired or restricted — check each status¶
| Item | Status | Replacement |
|---|---|---|
| AWS IoT FleetWise | Closed to new customers; existing customers may continue [1] |
Exclude from new-robot defaults — evidence |
| AWS RoboMaker | 🔴 Discontinued (2025-09-10) [1] |
EC2 G6e/G7e + Isaac Sim AMI + AWS Batch |
| SageMaker Edge Manager | 🔴 Discontinued (2024-04-26) [1] |
ONNX + IoT Greengrass V2 (+ SageMaker Neo) |
| IoT Greengrass V1 | 🔴 Discontinued (2026-06-01) [1] |
Greengrass V2 |
| Gazebo Classic 11 | 🔴 EOL (2025-01) [1] |
Gazebo Jetty/Harmonic |
| Trainium for VLA | ⚪ No public case [4] |
Currently CUDA/NVIDIA (state the risk when proposing) |
⚠️ Rumor watch (not true): "AWS IoT TwinMaker discontinued" is misinformation — TwinMaker is GA and open to new customers (low velocity). It is a third-party blog claim confused with SiteWise maintenance. Do not repeat. → pillar-3.
Promotion procedure (summary)¶
- Capture: collect candidates via a designated channel/emoji
- Filter: satisfy all inclusion conditions; separate release, evidence, use, and support.
- If it passes: the owning pillar's owner incorporates it via the standard template and removes it from the Radar
- If it falls short: keep it here as a one-liner, with the promotion condition stated
Full pipeline → maintenance.
owner: Youngjin · updated: 2026-09 · volatility: high (the Radar changes fast by nature — monthly review recommended)
-
World Foundation Model (WFM) — a large model trained to predict/generate the next scenes of the physical world. From text/video prompts it creates physically plausible video and scenarios to augment robot training data. 🎥 NVIDIA Cosmos introduction ↩
-
SysID (System Identification) — measuring the real robot's physical parameters (friction, mass, motor response) to calibrate the simulator to the real hardware. ↩
-
sim-to-real — transferring a policy trained in simulation to a real robot, or the methodology for doing so. The physical and visual differences between simulation and reality (the domain gap) mean a naive transfer collapses performance. 🎥 NVIDIA sim-to-real robotics showcase ↩
-
Physics engine — the core software of a simulator that numerically computes rigid-body dynamics, contact, friction, and collision. An engine's accuracy-speed trade-off drives the choice of simulator (Isaac/MuJoCo/Genesis). ↩
-
MCP (Model Context Protocol) — an open standard protocol connecting agents to tools and data sources. Often likened to "USB-C for agents"; experiments exposing robot skills as MCP servers are growing. ↩
-
ROS 2 (Robot Operating System 2) — the de facto standard open-source middleware for robot software. A distributed architecture in which sensor and control nodes communicate over topics; the shared foundation of industrial and research robot stacks. ↩
-
LLM agent — software in which a large language model plans on its own, selects and calls tools (APIs, robot skills), and carries out multi-step tasks. Unlike simple Q&A, the key point is that it "acts." ↩
-
Dexterity — a robot hand/arm's ability to manipulate objects with the precision and finesse of a human hand. Unlike a simple gripper's pick-and-place, it refers to contact-rich, complex manipulation such as reorienting an object or operating a tool with a five-fingered hand. ↩
-
UMI (Universal Manipulation Interface) — a way of recording robot demonstration data without a robot, using a handheld gripper (with a camera and force sensor) that a person holds to capture motion and contact force. Skipping the robot itself greatly cuts data-collection cost and safety overhead. ↩