AI-curated news on the latest developments in humanoid robotics
Newest humanoid robots announced & released
Agility Robotics unveiled upgrades to its Digit logistics humanoid including natural language command processing powered by a local LLM. The new vision system doubles object recognition speed and now handles partial occlusion. Amazon has expanded its Digit pilot to 12 fulfillment centers.
Chinese robotics company Unitree Robotics launched the G1 general-purpose humanoid at a breakthrough price point of $16,000, dramatically undercutting competitors. The G1 stands 127 cm tall, weighs 35 kg, and features 43 degrees of freedom with whole-body force sensing. Early orders have come from universities in 28 countries.
Norwegian robotics startup 1X Technologies launched a controlled beta testing program for its NEO home assistant humanoid robot. Fifty families in Oslo are hosting NEO robots for 6-month evaluations covering everyday household tasks and safety protocols. The company reports zero safety incidents across 12,000+ operational hours.
Apptronik unveiled an industrial-hardened variant of its Apollo humanoid robot designed for outdoor construction environments. The new model features IP65 dust and water resistance, reinforced joints rated for 50,000 cycles, and can operate in temperatures from -10°C to 55°C.
Sanctuary AI announced that its latest Phoenix humanoid matched human-level performance on 50 standardized manipulation benchmarks. The achievement follows 18 months of training using the company's Carbon AI control system. Phoenix can now generalize learned skills to novel objects without retraining.
UBTECH Robotics deployed its Walker X humanoid in a Shanghai hospital for patient assistance and logistics tasks. Three Walker X units operate on the neurological ward, delivering medications and guiding patients between departments. The hospital reports a 23% reduction in nursing staff time spent on logistics tasks.
Fourier Intelligence demonstrated its GR-1 therapy-focused humanoid assisting patients in lower-limb rehabilitation exercises. The robot adapts resistance and range of motion in real time based on patient feedback. Clinical trials involving 40 patients showed 31% faster recovery metrics compared to standard care.
Companies, funding & partnerships
Goldman Sachs Research revised its humanoid robot market size forecast upward to $38 billion by 2035, tripling its previous estimate. The upgrade reflects accelerating commercial deployments and faster-than-expected unit economics improvements. Manufacturing, logistics, and elder care are identified as the three highest-growth sectors.
Austin-based Apptronik raised $350 million in Series A funding led by Google Ventures with participation from existing DARPA program partners. The funding follows a successful 6-month DARPA performance evaluation of the Apollo humanoid. CEO Jeff Cardenas announced plans to build a 200,000 sq ft robot manufacturing facility in Texas.
SoftBank Robotics announced a complete redesign of its Pepper social robot with a large language model-first control architecture. The new Pepper features real-time conversation powered by a custom 70B parameter model running on NVIDIA Jetson hardware. Commercial deployments target airports and retail environments starting Q3 2025.
Foxconn Technology Group signed an agreement with Nvidia to deploy 10,000 Nvidia-powered humanoid robots across its Taiwan factories over three years. The robots will handle precision assembly tasks for consumer electronics where consistent quality is critical. Foxconn expects the program to reduce manufacturing defect rates by up to 18%.
The South Korean Ministry of Science and ICT announced a 1.4 trillion won ($1B) national investment to develop globally competitive humanoid robots by 2030. The 5-year program will fund 12 domestic robotics companies and 30 university research laboratories. Hyundai Robotics and Samsung have confirmed participation.
Amazon Robotics completed the acquisition of Covariant AI, whose robot foundation model powers autonomous manipulation in over 50 robot platforms worldwide. The deal, valued at approximately $750 million, brings 200 AI researchers focused on general robot intelligence into Amazon. Covariant's RFM-1 model will be integrated across Amazon's fulfillment robot fleet.
Figure AI secured a $675 million Series B funding round led by Microsoft and joined by OpenAI, Amazon, and Nvidia. The investment will accelerate production of the Figure 02 humanoid robot and expand the AI research team to 500 engineers. CEO Brett Adcock confirmed commercial sales are on track for 2025.
AI & engineering advances powering humanoids
Google DeepMind published RT-X, a generalist robot manipulation policy trained across 22 different robot platforms. The cross-embodiment training approach improves generalization dramatically, enabling robots to successfully complete tasks they were never directly trained on. Model weights have been released for research use.
A Stanford robotics team demonstrated 95% sim-to-real policy transfer using "adaptive physics perturbation" domain randomization. The method systematically varies simulation physics parameters to build robustness, resulting in policies that work on real hardware without physical robot training. Robot training time was reduced by 80%.
A joint team from ETH Zurich and Carnegie Mellon University unveiled an underactuated hand design that achieves 96% of human dexterity scores on standardized manipulation tests. The hand uses only 4 motors to control 20 degrees of freedom through clever tendon routing. Manufacturing cost is estimated at under $1,000.
UC Berkeley researchers demonstrated that current large language models can translate natural language task descriptions into executable robot motion primitives with zero task-specific training. The system completed 73% of novel household manipulation tasks described purely in text. This signals a shift toward general-purpose robot programming via language.
A collaboration between QuantumScape and Boston Dynamics developed solid-state battery packs optimized for humanoid robot power profiles. The technology delivers 4× the energy density of lithium-ion packs, extending single-charge operation from 2 to 8 hours. Field trials with Atlas robots confirmed the performance gains under real workloads.
Carnegie Mellon University researchers developed a low-cost tactile sensing skin detecting forces below 0.1 Newtons with millimeter spatial resolution. The sensor array costs under $50 per hand to manufacture and enables robots to handle eggs and glassware without breaking them. Three commercial robot companies have licensed the technology.
Stanford researchers developed a video-language-action model that extracts robot skill primitives directly from YouTube instructional videos without robot action labels. The model successfully learned 200 household manipulation skills from internet video, requiring only 30 minutes of real robot demonstration per skill. This drastically reduces data collection burden.
Deployed humanoids in action
An Amazon pilot facility in Seattle processes over 10,000 package sorting operations per day using 20 Agility Digit humanoid robots. The deployment reduced conveyor belt dependency by 35% and handles packages up to 11 kg. Amazon confirmed a $1 billion commitment to expanding the Digit program across its fulfillment network.
A major hospital system in Osaka deployed 15 humanoid robots to assist elderly patients with mobility, medication reminders, and light physical therapy guidance. Robots operate during night hours when nursing staff is reduced, handling 40% of routine patient check interactions. Patient satisfaction scores for overnight shifts improved by 28%.
US construction company Mortenson completed a 6-month pilot using Atlas-based humanoid robots for rebar placement and structural inspection at a high-rise construction site. The robots successfully completed tasks in environments unsafe for human workers including elevated platforms and confined spaces. The firm estimates 15% labor cost reduction for specific high-risk tasks.
A major European supermarket chain deployed humanoid robots in 25 stores for overnight shelf restocking operations. The robots successfully stock 850 items per shift and handle 92% of standard SKUs without errors. The deployment allows human staff to focus on customer service during peak hours. Plans to expand to 200 stores are underway.
The US Department of Energy approved humanoid robots for routine inspection tasks inside high-radiation zones at two nuclear facilities. The robots perform valve checks, sensor readings, and visual inspections — eliminating human radiation exposure for these tasks entirely. Annual dose savings of 340 person-rem per facility are projected.
1X Technologies concluded a 6-month home trial of its NEO robot with 50 elderly participants living independently. The robot successfully assisted with meal preparation, laundry folding, medication management, and light cleaning in 87% of requested tasks. Participants reported 4.2/5 satisfaction scores and a 34% reduction in required caregiver visits.
SpaceX is evaluating humanoid robotics systems for complex engine installation tasks in its Starship manufacturing facility at Boca Chica. Initial trials focus on high-torque fastener driving and cable routing in confined areas. SpaceX engineering teams developed a custom robot performance benchmark based on actual Raptor engine assembly steps.
Weekly update • LLMs, neural networks, vector search & scalable algorithms
A new distributed training framework reduces cross-device communication overhead during large neural network training by optimizing how gradient matrices are sharded and synchronized, cutting total training wall-clock time by roughly 35% on multi-node GPU clusters.
Researchers unveiled a sparsely-activated mixture-of-experts language model that matches the benchmark performance of dense models 3× its active parameter count. Dynamic expert routing cuts training compute significantly while preserving reasoning accuracy on standard evaluation suites.
A new sparse attention mechanism reduces the quadratic memory footprint of transformer neural networks at long context lengths by approximately 60%, with negligible loss in downstream task accuracy. The technique is already being integrated into several open-source LLM inference stacks.
A new graph-based approximate nearest-neighbor index improves vector search recall-latency tradeoffs, delivering 40% faster query scores on billion-vector embedding datasets used for retrieval-augmented generation. The index is designed to scale horizontally across commodity hardware.
A refined low-rank matrix adaptation method reduces the number of trainable parameters needed to fine-tune large neural networks, cutting fine-tuning time by roughly 5× while matching full fine-tuning accuracy on downstream benchmarks. The approach builds on established LoRA-style decomposition of weight matrices.
Weekly update • Qubits, error correction & quantum algorithms
Several major cloud providers began rolling out lattice-based post-quantum cryptographic key exchange by default across encrypted traffic, citing the long-term risk that sufficiently powerful quantum computers could eventually break current public-key encryption. The migration is expected to take several years to complete across dependent systems.
A new superconducting quantum processor surpassed 1,000 physical qubits while maintaining coherence times competitive with smaller predecessor chips. Improved chip packaging and reduced crosstalk between neighboring qubits were credited for the gains, bringing large-scale fault-tolerant designs closer to practical viability.
Researchers demonstrated a logical qubit encoded across dozens of physical qubits with an error rate below the surface-code threshold needed for error correction to improve reliability as systems scale. The result is viewed as a key milestone on the path toward fault-tolerant quantum computing.
A neutral-atom quantum computing platform showed that optically trapped atoms can be dynamically rearranged mid-circuit to create flexible qubit connectivity, sidestepping a major limitation of fixed-wiring superconducting designs. The system executed benchmark circuits using several hundred atomic qubits.
A refined variational quantum eigensolver approach reduced the circuit depth needed to simulate small molecule ground-state energies, cutting the gate count required on near-term hardware. Early tests on pharmaceutical candidate molecules showed results matching classical reference calculations.
Get daily humanoid robot updates delivered to your inbox
Extended weekly • week 6 of this synthesis cycle
Artificial intelligence and quantum computing have spent most of their histories advancing along separate tracks — one bound by data and parameter counts, the other by qubits and coherence times. That separation is narrowing. Machine learning models are being used to design better quantum error-correcting codes, while quantum-inspired sampling and optimization techniques are quietly finding their way into classical AI training pipelines. This weekly synthesis tracks where the two fields are actually touching, not just where they are mentioned in the same press release.
Investment overlap is becoming harder to ignore. Several of the same venture and sovereign funds backing frontier AI labs have also taken significant positions in quantum hardware startups over the past two years, and chip designers building AI accelerators are increasingly the same organizations exploring quantum co-processor architectures.
Researcher migration tells a similar story. Crossover publications — authors with both a machine learning and a quantum information background — have grown as a share of papers submitted to major AI and physics venues, and a number of senior hires this year have moved directly between AI labs and quantum hardware teams.
None of this should be mistaken for quantum computers becoming useful AI accelerators anytime soon. Today's noisy intermediate-scale devices remain error-prone, hard to reproduce results on outside a handful of labs, and nowhere close to the throughput classical GPU clusters offer for training large models.
The more immediate and less speculative connection is defensive: cryptographic infrastructure underpinning cloud AI systems is being migrated toward post-quantum standards now, years before large-scale fault-tolerant quantum computers are expected, precisely because the AI industry's dependence on encrypted data at rest and in transit makes it a high-value target for "harvest now, decrypt later" attacks.
Fault tolerance remains the milestone the whole field is waiting on — a quantum computer that can run arbitrarily long computations without errors accumulating faster than they can be corrected. What has changed this year is the pace at which AI-assisted design work is compressing the engineering timeline toward that milestone, even without altering the underlying physics.
Some of the most concrete crossover work is happening in error correction. Designing codes that protect logical qubits from noise is a brutal combinatorial search problem, and several hardware teams now use reinforcement-learning agents to discover decoding strategies and code layouts that outperform hand-designed alternatives. The qubits stay firmly quantum; the discovery process increasingly is not.
The traffic runs the other way too. Quantum-inspired sampling methods — algorithms that borrow the mathematics of quantum annealing without needing actual quantum hardware — are being tested as faster alternatives to standard gradient-based optimizers for certain classical machine learning training problems, particularly in combinatorial and scheduling-heavy domains.
Calibrating and operating a quantum processor is itself a machine learning problem in disguise. Pulse-level control, readout-error mitigation, and compensating for slow drift in qubit frequencies increasingly rely on models trained on the device's own telemetry, replacing manual tuning that used to consume a large share of lab time.
The honest summary is that AI and quantum computing are not merging into one technology — they are becoming better tools for building each other. Quantum research supplies AI with new problems, new datasets, and new security constraints; AI supplies quantum research with faster calibration, better error correction, and a shorter path from prototype to reliable hardware. Next week's edition will track how far that loop has turned.
