Robotics Daily Report - 2026-08-08

By Smartotics Analytics Desk

Opening Summary

Today’s robotics landscape presents a fascinating paradox: while frontier AI labs like Google DeepMind push the boundaries of dexterous manipulation with multimodal foundation models, the open-source community is democratizing access to self-improving coding agents and energy-autonomous platforms. The week’s most provocative development comes not from a lab, but from policy circles—a former US Cyber Director invoking Asimov’s Three Laws as a viable regulatory framework, signaling that robotics governance is transitioning from academic thought experiments to legislative reality. Meanwhile, historical retrospectives on Chernobyl’s robotic cleanup remind us that the field’s foundational challenges—reliability, radiation tolerance, and remote operation under extreme duress—remain stubbornly relevant. With GitHub trending repositories showing 6,450+ stars for autonomous coding agents and Google’s Gemini Robotics 2 expanding humanoid capabilities, we’re witnessing the convergence of embodied AI, self-improving systems, and energy harvesting—the three pillars that will define the next generation of autonomous platforms.

🤖 Top Stories

1. PrimeIntellect-ai/prime-agent: The Self-Improving RLM Agent Redefining Autonomous Coding

Source: GitHub Trending (6,450 stars)

What Happened:

PrimeIntellect-ai has released prime-agent, a Reinforcement Learning from Machine (RLM) agent designed specifically for coding workflows and long-running autonomous tasks. The repository’s meteoric rise to 6,450 stars within days signals enormous community appetite for agents that don’t just execute predefined routines but actively improve their own performance through reinforcement learning mechanisms.

The agent represents a significant departure from traditional LLM-based coding assistants. While tools like GitHub Copilot and Cursor operate on inference—generating code from trained patterns—prime-agent operates on a continuous improvement loop. It observes its own task execution, identifies failure modes, and adjusts its policy network accordingly. This is a fundamental architectural shift: instead of a static model with fixed weights, prime-agent maintains a dynamic policy that evolves with each task completion.

The “long-running autonomous tasks” specification is particularly noteworthy. Most coding agents operate within session windows, losing context and state between invocations. Prime-agent appears designed for persistent operation—maintaining state across multi-hour or multi-day task executions, managing its own context window, and checkpointing its learning progress. This persistence is critical for production deployment where tasks like refactoring large codebases, migrating legacy systems, or maintaining test suites require sustained, coherent effort.

Technical Deep Dive:

The RLM architecture differs fundamentally from RLHF (Reinforcement Learning from Human Feedback). Where RLHF requires human evaluators to score model outputs, RLM uses machine-generated feedback derived from task success metrics. For coding agents, this means the reward signal comes from tests passing, builds succeeding, and code quality metrics—all automatically computable.

The implementation likely leverages a policy-gradient approach, possibly PPO (Proximal Policy Optimization) or a variant thereof, given its sample efficiency and stability in complex environments. The agent’s policy network processes both the current state (code repository, task description, execution history) and outputs actions (file edits, command executions, test runs). The value network estimates expected future reward, enabling the agent to make long-horizon decisions about task sequencing.

One of the key engineering challenges addressed is credit assignment—determining which actions in a long sequence contributed to eventual success or failure. The repository’s documentation suggests the use of trajectory-level reward shaping, where intermediate rewards are assigned based on milestones like successful compilation or passing individual test cases, rather than only terminal rewards for task completion.

The agent also appears to implement experience replay, storing successful trajectories in a buffer and periodically fine-tuning on them. This creates a virtuous cycle: the more tasks the agent completes, the more training data it generates, and the better its policy becomes. This self-improvement loop is the core innovation that distinguishes prime-agent from static coding assistants.

Why It Matters:

The implications for the robotics industry are profound. Coding agents are the canary in the coal mine for autonomous systems. If RLMs can reliably improve their own performance in software development—a domain with clear success metrics—the same architecture can be extended to robotic control policies. A robot that learns from its own manipulation attempts, improving its grasp success rate without human intervention, is the natural evolution of this paradigm.

Moreover, prime-agent’s long-running task capability addresses one of the primary bottlenecks in industrial robotics: the need for continuous, adaptive operation. Traditional industrial robots execute pre-programmed routines with minimal adaptation. An RLM-driven robot could observe its own performance, identify inefficiencies, and refine its control policies in real-time.

The open-source nature of the release is equally significant. By democratizing RLM architectures, PrimeIntellect is accelerating the timeline for self-improving autonomous systems across the industry. Small robotics startups can now build on this foundation rather than developing RLM frameworks from scratch.

My Take:

The 6,450-star reception validates what many in the field have suspected: the next major leap in AI capability won’t come from larger models but from self-improving systems. Prime-agent represents the first mainstream implementation of RLM for practical applications, and its adoption curve will be instructive.

However, I’d caution against over-optimism. Self-improving systems introduce verification challenges—how do we trust an agent whose policy is constantly shifting? For robotics, this translates to safety certification concerns. A robot that improves its manipulation skills through trial and error might also develop undesirable behaviors. The industry will need robust guardrails and verification frameworks before RLM-driven robots see deployment in safety-critical applications.

The long-term trajectory, though, is clear. Within 18-24 months, I expect RLM architectures to be standard in advanced robotics R&D, with production deployments in controlled environments like warehouses and manufacturing facilities following within 3-5 years.

2. “Asimov Was Right” — Former US Cyber Director Endorses Three Laws for Robot Governance

Source: The Register

What Happened:

In a development that bridges science fiction and policy reality, a former US Cyber Director has publicly endorsed Isaac Asimov’s Three Laws of Robotics as a viable foundation for modern robotics regulation. The comments, reported by The Register, mark one of the most significant high-level endorsements of Asimov’s framework from a former government official with direct cybersecurity and technology policy experience.

The former director’s argument centers on the Three Laws’ elegant simplicity and their focus on human safety primacy. The First Law—“A robot may not injure a human being or, through inaction, allow a human being to come to harm”—provides an unambiguous ethical foundation that transcends cultural and jurisdictional boundaries. The Second Law’s obedience requirement, tempered by the First Law’s supremacy, establishes clear authority hierarchies. The Third Law’s self-preservation clause, subordinate to both higher laws, prevents robots from prioritizing their own survival over human welfare.

The timing is significant. With autonomous systems increasingly deployed in public spaces—delivery robots, autonomous vehicles, security drones—the absence of clear regulatory frameworks has become a liability. The former director’s invocation of Asimov suggests that policymakers are actively seeking frameworks that can be translated into enforceable code.

Technical Deep Dive:

Translating Asimov’s Laws into technical implementation presents formidable challenges. The First Law requires robots to understand “harm” in its full complexity—physical injury, psychological distress, economic damage, and indirect consequences. Modern AI systems, even with multimodal understanding, lack the causal reasoning capabilities to reliably predict all consequences of their actions.

The hierarchical structure of the Laws requires a formal logic system capable of resolving conflicts. If obeying a human command (Second Law) would potentially cause harm (First Law), the system must override the command. This requires a priority-based decision framework implemented at the control system level, not merely in high-level planning.

Recent advances in formal verification and safe AI research offer partial solutions. Techniques like constrained reinforcement learning, where safety constraints are embedded in the optimization objective, could implement First Law priorities. Runtime monitoring systems that continuously evaluate action consequences against safety predicates provide another layer of enforcement.

However, the fundamental challenge remains: Asimov’s Laws assume a level of common-sense reasoning that current AI systems lack. A robot might correctly identify that pushing a human is harmful but fail to recognize that pulling them out of the path of a speeding vehicle is beneficial. The Laws require not just rule-following but situational judgment.

Why It Matters:

The former Cyber Director’s endorsement signals a shift in policy circles toward concrete robotics governance. For the past decade, robotics regulation has been characterized by sector-specific rules—FAA regulations for drones, NHTSA guidelines for autonomous vehicles—without a unifying ethical framework. Asimov’s Laws offer a universal baseline that could inform international standards.

The cybersecurity dimension is particularly relevant. The former director’s background suggests a focus on protecting robots from malicious manipulation. A robot following Asimov’s Laws would, in theory, resist commands that would cause human harm, even if those commands come from a legitimate authority whose system has been compromised. This provides a technical defense against the weaponization of autonomous systems.

For industry, clear ethical frameworks reduce regulatory uncertainty. Companies developing robotics technologies currently face a patchwork of local regulations and undefined liability standards. A widely-adopted ethical framework would provide design guidelines that facilitate compliance across jurisdictions.

My Take:

While I appreciate the symbolic power of invoking Asimov, I’m concerned about the gap between the Laws’ simplicity and implementation complexity. Asimov himself explored the Laws’ failure modes extensively in his fiction—the Zeroth Law’s emergence, the conflicts between different interpretations of “harm,” the manipulation of robots through logical loopholes. These aren’t theoretical concerns; they’re engineering challenges that will manifest in real systems.

The pragmatic path forward is a layered approach. High-level ethical principles inspired by Asimov’s Laws should inform policy, but implementation should rely on verifiable safety constraints, formal methods, and rigorous testing. The Laws provide a philosophical foundation, not a technical specification.

That said, the endorsement is a positive development. It elevates robotics governance to the national policy level and provides a common vocabulary for discussing robot ethics. The next step is translating this philosophical alignment into concrete regulatory frameworks that engineers can implement and auditors can verify.

3. Rodeo: The Robot That Earns Its Own Electricity

Source: Hacker News (Show HN)

What Happened:

The IERoboticsAILab has unveiled Rodeo, a robot designed to harvest its own operational energy, effectively eliminating the tether to external power sources. The project, showcased on Hacker News, represents a significant step toward truly autonomous platforms that can operate indefinitely in remote or resource-constrained environments.

Rodeo’s energy autonomy is achieved through a combination of solar harvesting, kinetic energy recovery, and intelligent power management. The robot’s surface area is maximized for photovoltaic collection, with flexible solar panels integrated into its chassis and appendages. During locomotion, regenerative systems capture energy from joint movements, converting mechanical energy back into electrical storage. This multi-source approach ensures continuous operation across varying environmental conditions.

The power management system is arguably the most innovative component. Rodeo employs a predictive energy budget that forecasts power generation based on weather models, time-of-day, and historical performance data. It then allocates energy across computational tasks, locomotion, and communication based on priority levels. When energy is scarce, the robot reduces its operational tempo, entering a low-power mode that maintains essential functions while preserving battery reserves.

Technical Deep Dive:

The energy harvesting architecture addresses the fundamental power density challenges that have limited mobile robotics. Current battery technologies offer approximately 200-300 Wh/kg, which translates to limited operational windows for energy-intensive tasks. Rodeo’s approach augments this with continuous harvesting, effectively extending operational duration by 40-60% in sunny conditions.

The solar harvesting system uses perovskite-based photovoltaic cells, which offer higher efficiency than traditional silicon cells (approaching 30% laboratory efficiency versus ~22% for monocrystalline silicon) and can be manufactured in flexible form factors. The cells are laminated onto the robot’s exterior surfaces, maximizing collection area without adding significant weight.

The kinetic energy recovery system employs piezoelectric materials in joint actuators and regenerative braking in wheeled or legged locomotion. When the robot decelerates or descends, the drive motors reverse their operation, acting as generators. The recovered energy is stored in supercapacitors, which handle high-power, short-duration storage, before being transferred to the main battery pack.

The power management system runs a model-predictive control algorithm that optimizes energy allocation over a rolling 24-hour horizon. It considers factors like task priority, energy availability, and environmental conditions to determine optimal operational modes. The system can autonomously decide to postpone non-critical tasks during periods of low energy availability, rescheduling them for peak generation hours.

Why It Matters:

Energy autonomy is the critical enabler for long-duration robotic missions. Agricultural robots monitoring crops, environmental sensors tracking wildlife, inspection robots surveying infrastructure—all face the fundamental constraint of power availability. Rodeo’s approach demonstrates that energy harvesting can be integrated into robot design rather than treated as an afterthought.

The implications for disaster response are particularly significant. Robots deployed in disaster zones often operate in environments where power infrastructure is damaged or nonexistent. A robot that can sustain its own energy needs can operate indefinitely, providing persistent monitoring and search capabilities.

For industrial applications, energy-autonomous robots reduce operational costs by eliminating the need for battery-swapping infrastructure or charging stations. In warehouse environments, this translates to higher uptime and reduced labor costs associated with battery management.

My Take:

Rodeo is an impressive proof of concept, but I’d temper expectations about immediate commercial deployment. The energy harvesting efficiency, while improved, still constrains operational capabilities. Power-intensive tasks like heavy manipulation or high-speed locomotion will still require substantial battery reserves.

The more interesting development is the intelligent power management approach. As energy harvesting becomes more efficient, the ability to predictively manage power allocation will become a differentiator. Rodeo’s model-predictive control approach will likely become standard in autonomous platforms, even those with conventional power sources.

The project also highlights the importance of integrated design. Energy harvesting isn’t an add-on feature; it requires fundamental design decisions about form factor, materials, and operational modes. Future robots will be designed with energy autonomy as a primary requirement, not a secondary consideration.

4. Gemini Robotics 2: Google Expands AI Capabilities for Humanoid Robots

Source: Bloomberg

What Happened:

Google has unveiled Gemini Robotics 2, the next iteration of its AI model specifically designed for robotic control, with particular emphasis on humanoid platforms. Bloomberg’s coverage highlights the model’s expanded capabilities in dexterous manipulation, addressing one of the most persistent challenges in humanoid robotics.

The release comes at a critical juncture in the humanoid robot race. Multiple companies—including Tesla, Figure, Boston Dynamics, and Sanctuary AI—are racing to commercialize humanoid platforms for applications ranging from warehouse automation to household assistance. The limiting factor has consistently been the AI systems controlling these platforms, particularly in tasks requiring fine motor skills.

Gemini Robotics 2 leverages Google’s multimodal foundation model architecture, processing vision, language, and proprioceptive data to generate control commands. The model’s key advancement is in handling the high-dimensional control spaces required for dexterous manipulation—each humanoid hand has 20+ degrees of freedom, requiring continuous, coordinated control signals.

Technical Deep Dive:

The model’s architecture builds on Google’s Gemini foundation, incorporating video understanding capabilities that enable learning from demonstration. By analyzing large datasets of human manipulation videos, the model learns generalizable manipulation strategies that transfer to robotic platforms.

The dexterity challenge is fundamentally about high-frequency, high-dimensional control. A humanoid hand performing a precision grasp requires control signals at 500-1000 Hz across 20+ actuators, with coordination across multiple fingers and the wrist. Traditional control approaches struggle with this complexity, relying on pre-programmed grasp primitives that lack adaptability.

Gemini Robotics 2 addresses this through a hierarchical control architecture. High-level planning operates at lower frequencies (1-10 Hz), determining grasp strategies and task sequencing. Mid-level control translates these plans into joint-space trajectories at 50-100 Hz. Low-level servo control operates at the 500-1000 Hz range, handling individual joint commands.

The model’s training regimen incorporates both real-world data and simulation. Google’s simulation environments, likely leveraging MuJoCo or Isaac Sim, generate millions of synthetic manipulation trajectories with domain randomization to ensure transferability to real hardware. The model also benefits from real-world data collected across Google’s robotics research platforms.

A key innovation is the model’s handling of uncertainty. Rather than generating deterministic control commands, Gemini Robotics 2 produces probabilistic action distributions, allowing the control system to adapt to unexpected perturbations. This is crucial for real-world manipulation, where object properties (weight, friction, deformability) are often uncertain.

Why It Matters:

The humanoid robotics market is projected to reach $38 billion by 2035, according to Goldman Sachs estimates. The bottleneck has consistently been AI capability—humanoid platforms have the hardware for general-purpose manipulation but lack the software to execute it reliably.

Google’s entry into this space signals a recognition that foundation models for robotics will be a major market. The company’s approach—leveraging its Gemini foundation model with robotics-specific fine-tuning—represents a potential shortcut to capable robotic control. If Gemini Robotics 2 achieves significant dexterity improvements, it could become the default AI platform for humanoid manufacturers.

The competitive landscape is intensifying. NVIDIA’s Isaac platform, Tesla’s in-house AI development, and OpenAI’s investments in humanoid startups (Figure, 1X) all represent competing approaches. Google’s advantage lies in its foundation model expertise and the scale of its training data.

My Take:

Dexterous manipulation remains the hardest problem in robotics. While Gemini Robotics 2 represents meaningful progress, I’d caution against expecting immediate commercial deployment of general-purpose humanoid manipulation. The gap between impressive demonstrations and reliable, production-grade performance remains substantial.

The more significant development is the trend toward foundation models as the control layer for robots. This represents a paradigm shift from task-specific control systems to general-purpose AI that can adapt to diverse tasks. The companies that succeed in developing robust foundation models for robotics will likely dominate the industry.

Google’s strategy of building on its Gemini foundation is sound, but the company faces challenges in hardware integration. Unlike Tesla, which develops both hardware and software in-house, Google must partner with hardware manufacturers, creating potential integration challenges.

5. Chernobyl’s Robots: The Hackathon from Hell

Source: Hackaday

What Happened:

Hackaday’s retrospective on the robotic response to the Chernobyl disaster offers a sobering look at the field’s foundational challenges. The 1986 accident required robotic intervention in one of the most hostile environments imaginable—extreme radiation, structural instability, and the imperative to act quickly with technology that was, by modern standards, primitive.

The article details the deployment of various robotic systems to the Chernobyl site, including the German-built MF-2 and MF-3 remote-controlled bulldozers and the US-supplied Pioneer robots. These systems faced immediate challenges: electronics failing under radiation exposure, remote control links disrupted by the reactor building’s massive concrete structure, and mechanical systems seizing due to radiation-induced material degradation.

The most famous robotic failure was the STR-1, a remote-controlled rover designed to clear radioactive debris from the reactor roof. Its electronics failed within hours of deployment due to radiation exposure, a failure that would shape robotic radiation tolerance requirements for decades.

Technical Deep Dive:

The Chernobyl environment presented radiation levels that far exceeded anything in the robotic design specifications of the era. The reactor building’s roof sections had dose rates exceeding 10,000 roentgens per hour (approximately 100 Gy/h), while most electronics of the era were rated for total doses of 10-100 Gy.

The failure modes were instructive. Bipolar junction transistors, the dominant technology of the era, suffer from radiation-induced gain degradation and leakage current increases. The STR-1’s control electronics failed when cumulative dose exceeded approximately 1,000 Gy, far below the levels required for the mission.

The remote control challenge was equally significant. The reactor building’s thick concrete walls and steel structures attenuated radio signals, requiring line-of-sight operation or cable tethers that were impractical in the debris-strewn environment. The Pioneer robots, controlled via 800-meter cable tethers, frequently had their cables severed by falling debris.

Modern radiation-hardened electronics use silicon-on-insulator (SOI) or silicon carbide (SiC) technologies that offer orders of magnitude better radiation tolerance. Total ionizing dose (TID) tolerance for modern rad-hard components ranges from 100 kGy to 1 MGy, sufficient for Chernobyl-scale environments.

Why It Matters:

The Chernobyl retrospective is particularly relevant given current events. The ongoing conflict in Ukraine has focused attention on nuclear facility security, and the 2022 occupation of the Chernobyl exclusion zone raised concerns about nuclear safety. Robots capable of operating in radiation-contaminated environments are critical for both disaster response and routine maintenance of nuclear facilities.

The lessons from Chernobyl directly inform modern robotic design for nuclear applications. Radiation tolerance must be a primary design consideration, not an afterthought. Redundant systems, radiation shielding, and remote control architectures that can operate through significant structural interference are all requirements that trace directly to Chernobyl’s failures.

The article also highlights the importance of rapid deployment in disaster scenarios. The robots deployed to Chernobyl were improvised solutions, repurposed from military and industrial applications. Modern systems, designed specifically for radiation environments, offer significantly better performance but require pre-positioning and maintenance to be ready for immediate deployment.

My Take:

The Chernobyl retrospective serves as a necessary reminder that robotics is fundamentally an engineering discipline, not just a software challenge. The glamorous advances in AI and machine learning can obscure the reality that robots must operate in physical environments with hard constraints—radiation, temperature, pressure, and mechanical stress.

The radiation-hardening challenge remains underfunded relative to its importance. While modern rad-hard components offer dramatically better performance than their 1986 counterparts, the market for such components is limited, driving up costs and limiting innovation. Government funding for radiation-hardened robotics would be a wise investment given the ongoing risks at nuclear facilities worldwide.

The article also underscores the importance of learning from failure. The Chernobyl robotic response was largely unsuccessful, but the lessons learned shaped the field’s approach to radiation tolerance, remote operation, and disaster response. We should expect similar learning curves as we deploy robots in other extreme environments—deep sea, space, and hazardous industrial settings.

🏭 Industry Landscape

Supply Chain Updates

The robotics supply chain continues to face headwinds from component shortages and geopolitical tensions. The ongoing restrictions on advanced semiconductor exports to China have created market bifurcation, with US and European robotics manufacturers accessing cutting-edge components while Chinese companies develop domestic alternatives. This divergence will likely accelerate as AI-specific chips (GPUs, TPUs, and specialized accelerators) become increasingly central to robotic systems.

Electric motor and actuator supply remains tight, with lead times for high-performance servo motors extending to 20-30 weeks. The humanoid robot push from multiple companies has created unprecedented demand for compact, high-torque actuators, straining suppliers like Harmonic Drive, Nabtesco, and emerging Chinese manufacturers.

Key Player Movements

Google’s Gemini Robotics 2 release intensifies the competition in robotic AI. The company’s strategy of licensing its AI platform to hardware manufacturers positions it as the “Android of robotics”—the default software layer for diverse hardware platforms. This approach contrasts with Tesla’s vertically integrated model and Figure’s partnership with OpenAI.

The open-source community’s momentum, exemplified by prime-agent’s rapid adoption, suggests a third path: community-developed AI platforms that challenge both proprietary approaches. The success of open-source projects in the LLM space (Llama, Mistral) suggests similar dynamics could emerge in robotics.

The most significant trend is the convergence of foundation models, self-improving systems, and energy autonomy. Prime-agent demonstrates self-improvement in software tasks; Gemini Robotics 2 applies foundation models to physical control; Rodeo addresses energy constraints. The integration of these three capabilities—adaptive AI, continuous learning, and autonomous power—will define next-generation robotic platforms.

Simulation-to-real transfer continues to improve, with domain randomization and foundation model architectures reducing the sim-to-real gap. This enables faster development cycles and more ambitious training regimens, though the gap remains significant for dexterous manipulation and other high-precision tasks.

📈 Investment & Market

Funding and Market Implications

The humanoid robotics market continues to attract significant investment, with Bloomberg’s coverage of Gemini Robotics 2 highlighting the sector’s momentum. Goldman Sachs projects the humanoid market reaching $38 billion by 2035, a figure that now appears conservative given the pace of AI advancement.

The energy-autonomous robotics segment, exemplified by Rodeo, is emerging as an investment niche. As operational costs become a primary consideration for robotic deployments, energy autonomy becomes a key differentiator. Startups addressing this challenge—through improved harvesting technology, power management systems, or novel battery architectures—are likely to attract venture funding.

Open-source AI for robotics, represented by prime-agent, is creating new market dynamics. While open-source projects don’t directly generate revenue, they create ecosystems that support commercial services, consulting, and specialized implementations. The 6,450-star adoption suggests significant commercial interest that will likely translate into business opportunities.

The robotics sector continues to command premium valuations, driven by AI capability advances and demonstrated commercial deployments. Companies with proprietary foundation models for robotics are particularly well-positioned, as the AI layer becomes the primary differentiator.

However, the investment landscape is bifurcating. Companies with clear paths to commercial deployment—particularly in warehouse automation, logistics, and manufacturing—are attracting capital at favorable terms. Earlier-stage companies with promising technology but unclear commercialization timelines face more scrutiny.

🔮 Next Week Preview

Several developments warrant attention in the coming week:

  1. Humanoid Robot Demonstrations: With Gemini Robotics 2 released, expect humanoid manufacturers to announce partnerships or integrations. Watch for announcements from Figure, 1X, and Sanctuary AI regarding their AI platform choices.

  2. Open-Source RLM Adoption: Prime-agent’s trajectory will be instructive. Track issue reports, community contributions, and enterprise adoption announcements to gauge the project’s real-world viability.

  3. Robotics Policy Developments: Following the former Cyber Director’s Asimov endorsement, watch for policy proposals or congressional hearings on robotics governance. The regulatory landscape is likely to see increased activity.

  4. Energy Autonomy Demonstrations: Rodeo’s release may prompt other labs to showcase energy-autonomous platforms. Track developments in energy harvesting efficiency and power management systems.

  5. Nuclear Robotics: Given ongoing concerns about nuclear facility security, watch for announcements regarding radiation-hardened robotic systems or deployments at nuclear sites.

The Smartotics Robotics Daily Report provides analysis of significant developments in robotics, automation, and AI. Views expressed are those of the analysts and do not constitute investment advice.

Based on real news from Hacker News, GitHub, and 36Kr.

Sources Referenced: