Top 10 Robotics Stories — April 24 to May 1, 2026

Executive Summary

This week's robotics signal centers on a hard collision between commercial humanoid expectations and the physical realities of manufacturing — Tesla's public reset on Optimus production and Gen 3 timing — set against a research-side acceleration in what foundation models for embodiment can actually do. Physical Intelligence's pi 0.7 demonstrated that a 5B-parameter VLA, when trained with diverse multimodal prompting (language, metadata, and subgoal images), can be steered zero-shot onto unseen embodiments and chores like espresso-machine operation and laundry folding, matching RL-finetuned specialists (Physical Intelligence, arXiv 2604.15483). Two other architectural threads sharpened: STARRY framed manipulation as action-centric world modeling rather than direct-action prediction (arXiv 2604.26848), and BeTTER's diagnostic benchmark exposed how much of today's apparent embodied reasoning is in fact visuospatial heuristics (arXiv 2604.18000). Together these continue the paradigm proliferation already visible in WAV (latent trajectory inference) and AgiBot's ViLLA-with-Action-CoT — the field is no longer settled on direct-action VLAs as the default.

The hardware story is bifurcated. On the humanoid side, the gap between research demos and depreciable industrial assets remains stark: Optimus admitted essentially zero useful work in 2025 with Fremont production deferred to mid-2026 on a repurposed Model S/X line (Electrek), while Hannover Messe nonetheless featured a humanoid named Alpha working in Siemens-instrumented factory cells across Europe, MENA, and North America (HANNOVER MESSE). Wheeled-legged platforms pushed dynamic dodging into reflexive territory (arXiv 2604.23761), and a comprehensive review made the case that legged robotics needs to escape the stationary-ground assumption to operate on ships, aircraft, and moving vehicles (arXiv 2604.20990).

On the surgical and commercial frontier, Intuitive Surgical's Q1 raised 2026 da Vinci procedure growth guidance even after disclosing a March phishing-driven breach, underscoring that medical robotics revenue continues to compound on a consumables/procedures model that is largely orthogonal to humanoid hype (MedTech Dive). And the field paid its respects to its own history: ICRA 2026 awarded its Most Influential Paper to the 2008 Reciprocal Velocity Obstacles work that today underwrites multi-agent navigation in nearly every autonomous fleet (UMD CS). Two further papers — MOMO on multimodal physical/verbal/graphical skill teaching and a planning-with-active-perception scene-graph framework — flesh out the human-robot teaching surface and the symbolic-meets-perceptual planning stack (arXiv 2604.20468, arXiv 2604.26988). The week's overall through-line: foundation-model architectures continue to fragment productively, while industrial humanoids are entering the deeply unglamorous phase where line builds, takt times, and certifications matter more than demo videos.

1. Tesla Optimus Q1 2026 Reset: Fremont in Late Summer, Gen 3 Slips to Mid-Year

Tesla's Q1 2026 earnings call delivered the most candid Optimus update in the program's history, and it was a marked retreat from prior commitments. Elon Musk acknowledged that "essentially zero" useful work had been done by Optimus in 2025 and confirmed that the Gen 3 reveal — previously promised for Q1 2026 — has slipped to mid-2026, with initial Fremont production now targeted for late July or August on a converted Model S/Model X manufacturing line (Electrek). A second Optimus production facility at Giga Texas is now scheduled for summer 2027 to manufacture Gen 4 units, pushing volume manufacturing further out than analysts had modeled (Longbridge).

The strategic admission matters more than the schedule slippage. Repurposing a depreciated, fully-automated Model S/X line for humanoid assembly is the right capital decision — those tooling slots are running well below capacity — but it also signals that Tesla has not yet engineered a clean-sheet humanoid line, which is exactly the kind of process commitment that distinguishes Toyota- or Hyundai-grade manufacturing programs from prototype workshops. Musk's framing that Optimus would be "the biggest product of all time" remains, but the program is now tracking on the same multi-year hardware-iteration curve as Cybertruck rather than the software-velocity curve Tesla repeatedly invokes (Electrek).

For the broader humanoid sector, the reset has two implications. First, it widens the credibility gap between Tesla and competitors who are already shipping or piloting hardware in customer environments — Boston Dynamics Atlas at Hyundai, Figure at BMW Spartanburg, Agility's Digit at GXO, and Accenture-SAP-Vodafone's Duisburg warehouse pilot disclosed at Hannover Messe — making the "first to volume" race far less Tesla-determined than 2024 narratives implied. Second, the explicit Gen 3/Gen 4 staging signals that hand dexterity, actuator density, and battery integration on Gen 2 are still well short of what Optimus needs for general-purpose manipulation, consistent with the hand-redesign debates that surfaced through 2025.

2. Physical Intelligence pi 0.7: Steerable VLA With Zero-Shot Cross-Embodiment Skills

Physical Intelligence (PI) released pi 0.7, a 5B-parameter vision-language-action model that operationalizes a key insight: a single VLA can be made steerable across embodiments and tasks if it is trained with diverse multimodal prompting modalities — natural-language instructions, structured metadata, and subgoal images — rather than language alone (Physical Intelligence, arXiv 2604.15483). The model demonstrates zero-shot transfer to unseen embodiments and chores including laundry folding and operating a commercial espresso machine, and it matches the performance of RL-finetuned specialists on the same tasks without per-task RL post-training (TechBuzz).

Architecturally, pi 0.7 is in the direct-action VLA lineage that PI established with pi-zero and pi-zero-FAST, but the diversification of prompt modalities is what enables the steerability claim. Subgoal-image prompts let the policy condition on a target observation without a verbal description, which is critical for tasks where the desired end state is easier to demonstrate than describe (folding states, drink presentation, instrument arrangement). Metadata prompts allow per-deployment specialization — robot kinematics, gripper type, scene priors — without retraining. Combined, these turn the VLA from a fixed mapping into something closer to a conditional policy library that can be reconfigured at deployment time.

The release is significant because it answers, at least empirically, the criticism implicit in this week's BeTTER benchmark (Story 8): namely that direct-action VLAs do not really reason. PI's evidence is that with a sufficiently broad and diverse multimodal prompting distribution, the model generalizes to genuinely new embodiments and tasks at quality comparable to bespoke RL — which is exactly the capability one would want from a "robot foundation model." The open question is robustness under distribution shift in long-horizon tasks, where WAV-style latent trajectory inference and STARRY-style world models (Story 4) propose alternative inductive biases that may compound advantages over many-step manipulation.

3. Hannover Messe 2026: Humanoid "Alpha" Lives at the Siemens Booth

Hannover Messe 2026 (April 20-24) treated humanoid robots not as side-show curiosities but as central exhibits, with the most-discussed deployment being a humanoid called Alpha operating live workflows at the Siemens booth (HANNOVER MESSE, YouTube demo). Siemens framed the integration as a productized partnership — Alpha runs on Siemens MES/PLC infrastructure with the company's industrial AI stack handling perception and orchestration — and signaled that level-up technology partnerships are being negotiated across Europe, MENA, and North America for next-stage trials.

This sits alongside the Accenture-SAP-Vodafone Duisburg pilot disclosed at the same show in week prior, where a humanoid is treated as a first-class work resource inside SAP EWM rather than a bolted-on automation island. The pattern across both deployments is that the integration story has moved from "robot performs task" to "robot is a labor unit visible to the planning and execution systems that already run the factory" — which is the only way humanoid economics close, because dispatch, exception handling, and skill loading need to be solved at the MES layer, not inside the robot.

The Hannover Messe humanoid track also previewed a sharper European stance on humanoid certification, with discussions on whether ISO 10218-2-style collaborative robot safety frameworks need a humanoid-specific extension — particularly for bipedal balance failure modes that have no fenced-cell analog. Resolving that is a prerequisite for the kinds of mixed human-humanoid floor deployments Siemens demonstrated (HANNOVER MESSE).

4. STARRY: Spatial-Temporal Action-Centric World Modeling for Manipulation

STARRY (arXiv 2604.26848, April 29) proposes treating robotic manipulation as a world-modeling problem rather than a direct-action problem, predicting future video frames conditioned on candidate actions and using the predicted rollout to plan (arXiv, HTML version). The "action-centric" qualifier distinguishes it from generic video prediction models like Cosmos: STARRY's tokenization and attention masks are structured around action conditioning, so the model is trained to be sensitive to small action perturbations rather than to produce visually plausible but action-invariant futures.

The motivation is the same one driving WAV (last week's latent-trajectory model): direct-action policies have to extrapolate over long horizons in a fundamentally one-step-Markov way, which compounds error multiplicatively. World-model approaches push the horizon problem into the model's own forward-rollout dynamics, where the inductive biases of video generation (smooth motion, object permanence, contact discontinuities) act as a prior. STARRY's results show meaningful gains on long-horizon manipulation benchmarks where direct-action VLAs degrade most.

For practitioners, STARRY is one of the clearest signals this week that the field is moving toward hybrid architectures where a world model handles imagination and a smaller policy handles closed-loop control — analogous to MuZero in game-play agents. The open engineering question is real-time inference: video-token generation is expensive, and the published numbers are not yet at the latency required for, say, contact-rich assembly. But the capability ceiling is now substantially higher than it was for direct-action VLAs alone (arXiv html).

5. Wheeled-Legged Reflexive Obstacle Evasion at the Dynamic Limit

A team led by Zhao et al. (arXiv 2604.23761, April 26) published "Agility Unleashed: High-Dynamic Reflexive Obstacle Evasion for Wheel-Legged Robots," demonstrating dodging behaviors that operate at the dynamic limit of wheel-legged platforms — the kind of reflex that needs to fire in tens of milliseconds to clear an unexpected obstacle while preserving balance (arXiv html). The architecture is a tightly coupled perception-control stack: short-horizon predictive sensing feeding a learned reactive controller that directly modulates wheel torque, leg extension, and CoM trajectory.

Wheel-legged platforms — the morphology Tencent's Robotic X and ETH's Ascento helped popularize — have always been pitched as combining wheeled efficiency with legged terrain competence, but their failure mode under dynamic obstacles has been the worst of both worlds: too much momentum to stop like a quadruped, too constrained at the wheel for arbitrary footstep replanning. Zhao et al. show that with sufficiently aggressive controller gains and learned reflexive policies, the platform can in fact dodge at speeds where neither pure-wheel nor pure-leg solutions would work — momentum becomes an asset rather than a liability if the controller can shape it.

The broader context is that wheel-legged platforms are quietly winning in last-mile and logistics use cases (Honda's UNI-ONE, Segway's S-Pod successors, the Direct Drive Tech wheel-legged delivery line) where efficiency and curb-mounting matter more than full bipedal repertoire. Reflexive evasion at dynamic limits is what unblocks deploying these into uncontrolled pedestrian environments at higher speeds, which is the real productivity gain over a pure-quadruped delivery robot (arXiv html).

6. ICRA 2026 Most Influential Paper: Reciprocal Velocity Obstacles (2008)

ICRA 2026's Most Influential Paper Award (covering the 2004-2008 publication window) went to "Reciprocal Velocity Obstacles for Real-Time Multi-Agent Navigation" by Jur van den Berg, Ming Lin, and Dinesh Manocha — a recognition long overdue given that RVO and its descendants (ORCA, NH-RVO, Hybrid RVO) underwrite essentially every modern multi-agent navigation system that scales beyond a handful of agents (UMD CS).

The technical insight that earned the award was the reciprocal assumption: rather than each agent treating others as moving obstacles to avoid (which produces oscillatory deadlocks), each agent assumes the others are also navigating cooperatively and shares the avoidance burden symmetrically. This converts an intractable joint planning problem into a per-agent local optimization that is provably collision-free under the reciprocity assumption and scales linearly in agent count. It is the algorithmic substrate behind crowd simulators, drone swarm coordination layers, AMR fleet managers (the Anki/Cozmo team called RVO "the only reason warehouse fleets work"), and large parts of game-engine pathfinding.

The award is a useful counterweight to the prevailing narrative that multi-agent coordination is now an MARL problem. RVO and its successors provide the safety and liveness guarantees that learned multi-agent policies still struggle to match — and as this week's other ICRA discussions emphasized, the productive path forward is hybrid systems where a learned high-level policy proposes preferences and an RVO-class layer enforces feasibility. Reading the original paper is still arguably the highest-leverage way to understand modern multi-agent navigation (UMD CS).

7. Robot Planning with Active Perception via Scene-Graph Reasoning

A new ICRA-track paper (arXiv 2604.26988, April 28) on "Robot Planning and Situation Handling with Active Perception" reframes long-horizon planning under partial observability around scene-graph structured representations, where the robot reasons over an explicit graph of objects, relations, and uncertainties and chooses both physical and perceptual actions to make progress (arXiv html). Active perception — choosing where to look, what to inspect, and when to disambiguate — is treated as a first-class action category alongside manipulation primitives.

The contribution is the integration of scene-graph reasoning with an LLM-based planner that operates over the symbolic graph rather than over raw text or pixels. The LLM proposes plans and revisions, the scene graph maintains an updated belief over the world, and a small set of perceptual primitives (look-at, probe, query) feed evidence back into the graph. This decouples reasoning from perception in a way that makes the planner robust to perceptual occlusions and ambiguities that pure end-to-end VLAs handle poorly.

In a week dominated by direct-action and world-model VLAs, this paper is the strongest reminder that classical planning representations — scene graphs, predicate logic, partial-order plans — are not obsolete. The integration of these representations with LLMs and modern perception stacks is producing systems that can reason about long-horizon tasks at a level direct-action policies cannot reach, particularly when the task involves searching for occluded objects, recovering from execution failures, or dynamically replanning around new constraints (arXiv html).

8. BeTTER: Unmasking the Illusion of Embodied Reasoning in VLAs

BeTTER (Benchmarking Embodied True Reasoning, arXiv 2604.18000) is a diagnostic benchmark designed to separate genuine embodied reasoning from visuospatial heuristic exploitation in vision-language-action models (arXiv html). The authors construct task families where surface visual cues (object color, position, distractor count) are systematically dissociated from the reasoning required to solve the task, and show that current VLAs collapse to chance or near-chance performance once the heuristics are removed.

The headline finding is that the VLM-to-VLA bridge is leaky: pretrained VLMs bring strong visuospatial priors that help on standard benchmarks, but those priors do not transfer to the action policy in a way that enables genuine multi-step reasoning. When the benchmark forces the policy to reason — for example, to infer which object is "third from the left when arranged by weight" — performance drops sharply, indicating that the action head is not consuming the VLM's reasoning chain so much as a flattened embedding of visual features.

For builders, BeTTER's value is methodological. It establishes a reproducible framework for asking "is this VLA actually reasoning?" rather than "does this VLA perform well on tasks where reasoning would help?" — a distinction that the field has been sloppy about. For pi 0.7 and STARRY-class systems, BeTTER's failure modes are a target: a model that solves BeTTER tasks at human-comparable rates will likely also solve a much wider class of long-horizon manipulation tasks. Conversely, beating standard benchmarks while failing BeTTER should be read as evidence of heuristic overfitting rather than capability (arXiv html).

9. Survey: Legged Robotics in Non-Inertial Environments

A comprehensive survey by Wang et al. (arXiv 2604.20990, April 22) on "Legged Robotics in Non-Inertial Environments" reviews the surprisingly under-served problem domain of legged locomotion on supports that are themselves moving — ship decks, aircraft carrier flight decks, aerospace platforms, and ground vehicles (arXiv). The stationary-ground assumption underlies almost every published locomotion controller, and the survey makes the case that this is now the binding constraint for several high-value deployment domains.

The technical issues fall into three buckets. Modeling: the contact dynamics on a non-inertial support need to include the support's own acceleration as a state, which breaks standard MPC formulations that assume world-frame inertial dynamics. State estimation: IMU-based estimators saturate or alias when the platform is itself accelerating, requiring fusion with platform-IMU data or visual-inertial schemes designed for non-inertial frames. Control: balance criteria like ZMP and capture point need redefinition relative to the moving support, and learned policies trained in stationary-ground simulators do not zero-shot transfer.

The survey is timely because several of the most promising commercial niches for quadrupeds (Boston Dynamics Spot at sea, Anymal on offshore platforms, military quadrupeds operating from vehicle ramps) all violate the stationary-ground assumption. For builders, the survey lays out a research agenda — improved sim environments with non-inertial primitives, dataset releases of platform-mounted teleoperation, and benchmark tasks like "walk a transverse line on a pitching deck." This is one of those fields where progress is gated more by the simulator and dataset gap than by algorithm novelty (arXiv).

10. MOMO: Multimodal Skill Learning From Physical, Verbal, and Graphical Demonstrations

MOMO (arXiv 2604.20468, April 22) presents a framework for robot skill learning that seamlessly fuses physical demonstrations (kinesthetic teaching), verbal instructions (natural language), and graphical inputs (sketches, annotated images, gesture overlays) into a single trainable policy (arXiv html). The key engineering contribution is a multimodal tokenizer that aligns these different teaching modalities into a shared representation, allowing a human teacher to start a skill demonstration kinesthetically, refine it verbally ("a little higher"), and then add a graphical correction (drawing the desired contact point on a tablet image of the workspace).

This addresses a real bottleneck in deployed robotics: teaching new skills today requires either expert programming (slow), kinesthetic-only demonstration (limited expressiveness), or specialized hardware (teleoperation rigs that don't generalize). MOMO's promise is that a domain-expert worker — without robotics training — can teach a skill using whichever modality is most natural for the part of the skill being conveyed, and the policy adapts. Early results show competitive sample efficiency relative to single-modality demonstration baselines, with the key gain coming on tasks where physical demonstration is awkward but verbal-plus-graphical correction is fast.

The strategic relevance is that MOMO is exactly the kind of teaching layer that humanoid programs need to make commercial economics work. The cost model for industrial humanoid deployment is dominated by skill onboarding time per new SKU or workflow — if that drops from days of expert engineering to hours of operator teaching, the addressable workflow inventory explodes by orders of magnitude. This dovetails with this week's pi 0.7 multimodal prompting work: pi 0.7 conditions a foundation model on diverse prompts at inference, MOMO conditions skill learning on diverse demonstrations at training. Both point at multimodality as the core lever for deployable robot intelligence (arXiv html).