Vishay Precision Group (VPG) will showcase custom force and torque sensing solutions for humanoid robotics at RoboBusiness 2026. These sensors are crucial because humanoid robots require precise, real-time feedback for safe and smooth interaction with unpredictable environments. VPG designs custom strain gage-based sensors built directly into robot structures to measure forces, torques, and multi-axis data. Specific solutions include miniature tactile sensors for fingertips, multi-axis sensors for joints, and joint torque sensors for actuators. VPG also offers complete engineering support to help move robotic designs from concept to large-scale production.
Robot safety functions rely on inputs like distance readings or stop signals, which can be deliberately manipulated to cause unintended behavior. Research shows that adversarial inputs, such as altered images or audio, can change a robot's actions despite protective systems being in place. This manipulation challenges safety assessments that often only consider accidental faults, necessitating a focus on deliberate cyber attacks. Safety teams must investigate the source and integrity of all information used by safeguards to ensure they remain effective. Strengthening safety requires testing against manipulated inputs and continuously updating evidence throughout the robot's lifecycle.
Claire interviewed Michelle Lu of Vsim Technology about using simulation and artificial intelligence to teach robots new skills. Lu previously co-founded Isaac Gym, a framework for GPU-accelerated reinforcement learning in robotics. Vsim was established to address limitations in simulation-based robotics AI concerning efficiency, scalability, and the sim-to-real gap. The Robot Talk podcast explores the field of robotics, artificial intelligence, and autonomous machines. This episode focuses on advancing robot learning through simulation and AI methods.
Boston Dynamics has redesigned the Atlas robot's hand to prioritize real-world applicability over purely humanlike aesthetics. The new design focuses on achieving a balance of capability, ruggedness, manufacturability, and low cost. This involves simplifying the design, such as removing the pinky, and increasing the degrees of freedom from seven to thirteen. The new hand is engineered to be strong, capable of using tools, and easily mass-produced and repaired. This approach suggests that scaling humanoid robots will require making design trade-offs for practicality.
Galbot has priced its ET1 bipedal humanoid in three versions, starting at 79,000 yuan. The Standard version is 79,000 yuan, the Pro is 139,000 yuan, and the Flagship is 179,000 yuan. The Pro tier includes "watch-and-learn" capabilities, while the Flagship offers more onboard compute and development options. These tiers differentiate features such as interactive learning, dynamic reasoning, and the inclusion of dexterous hands. The pricing establishes the cost for different levels of performance and advanced functionality in the humanoid robot.
Elon Musk stated that Tesla halved the RAM allocation for AI5 to 72GB and reduced AI6's by one-third to 144GB to accommodate Optimus production. This change was implemented to reduce costs while keeping memory bandwidth unchanged. Musk expects these reductions to have a negligible impact on Optimus performance because memory bandwidth is the primary constraint. This adjustment is intended to improve component supply and lower costs for scaling humanoid robot production. The update reflects a supply constraint in achieving the company's humanoid ambitions.
Melvin Schwarz described nearly two hours of factory work from Nucleus, showing visible interventions. The demonstration featured a 60% autonomous and 40% teleoperated operation. Nucleus now regularly conducts four-to-six-hour shifts, combining autonomy with human assistance in industrial deployments. The company's approach involves supervised operation and interventions to gather training data before achieving full autonomy. This demonstration provides context for understanding the balance between human and robot involvement in industrial tasks.
CEO James Zhou toured UBTECH's factory, detailing production capacity for humanoid robots. The 14,000-square-meter facility can potentially produce a humanoid every eight to ten minutes at maximum capacity. Monthly output is estimated to reach approximately 1,500 units based on these figures. The factory incorporates human-robot collaboration for material handling, with robots managing packaging and transport alongside workers. Furthermore, an automated vertical warehouse can store 112 humanoid robots within a small footprint.
Boston Dynamics introduced new four-finger hands for Atlas, increasing the degrees of freedom from seven to thirteen. These hands feature direct actuation and tactile sensors to enable object manipulation and tool operation. The design prioritizes dexterity, strength, and manufacturability, resulting in a hand size comparable to a human hand. The four-finger configuration allows for in-hand reorientation and gripping of tools. Engineers deliberately omitted a fifth finger to manage cost and potential failure points.
The research addresses the issue of contact force integration noise in simulators, which creates divergence in audit-chain logs for LLM-driven robots during object handoffs. The authors propose bounded-fidelity sim-as-demo-stage, a pattern that suppresses contact physics within defined handoff envelopes while maintaining full dynamics elsewhere. This method uses a MuJoCo mocap primitive driven by a Python adapter to achieve this control over simulation fidelity. Experiments show that by controlling envelope-edge timing jitter, audit chain reproducibility is achieved with near-zero engineering cost. This pattern allows benchmark designers to ensure audit-chain stability in simulations without incurring high fidelity costs.
A path tracking controller based on Nonlinear Model Predictive Control (NMPC) is developed for guiding agricultural tractors along a predefined path. The NMPC objective function incorporates multiple segments of a piecewise-linear reference path. Methods are also presented for selecting the most viable reference segments from the full path. Field tests demonstrated that the NMPC solver converged quickly and achieved a mean absolute cross-track error of 6.1 cm while tracking the curved path.
The paper proposes a method for legible planning in arbitrary PDDL domains by extending prior research on planning legibility. It avoids the need to construct ad-hoc planners by focusing on planning with off-the-shelf planners. The method incorporates a second-order theory of mind to estimate the observer's perspective, which helps in understanding goal disambiguation. Legibility is shown to be a trade-off with plan efficiency, requiring a regularizing factor for optimal balance. This approach has potential applications in human-robot teaming scenarios.
HumanoidTTT is a framework designed for test-time capability reuse in continual humanoid control to improve efficiency. It introduces Selective Full-Motion Reuse to allow direct reuse of complete motions from certified entry states, bypassing fresh generation. Test-Time Capability Consolidation adapts which qualified capabilities are retained in a store based on subsequent deployment reuse feedback. Experiments showed zero unsafe accepts and a 16.4$\times$ speedup over fresh generation. This method also improved avoided generator calls by 13.2 per 200 requests compared to a frozen counterpart.
The paper proposes a unified actuation-space framework for real-time, safe motion generation in multi-segment tendon-driven continuum robots. It utilizes an energy-based variable-curvature model to account for non-uniform bending and provides analytical Jacobians for kinematic calculations. A multipoint Control Barrier Function-Quadratic Programming (CBF-QP) enforces backbone clearance and respects actuation limits to ensure collision avoidance. Experiments showed a 96% collision-free success rate in static scenarios and 100% in dynamic ones, demonstrating real-time performance of 6.66 ms.
DriftOPD is a teacher-free, rollout-free framework for sequence-level on-policy distillation of continuous Vision-Language-Action (VLA) action experts. It decomposes the sequence-level reverse Kullback-Leibler divergence into a chunk-level term and a future-potential term reflecting long-horizon effects. The framework optimizes these terms using a one-step drifting objective and a Q-function critic learned from offline demonstrations. This allows for sequence-level optimization using only offline data and one-step action generation. DriftOPD achieves performance comparable to multi-step teacher policies in various VLA architectures.
IndoorBEV is a lightweight framework for real-time LiDAR perception in indoor mobile robots by using a height-aware Bird's-Eye-View (BEV) representation. It mitigates the loss of vertical geometric information by summarizing vertical point distribution using statistical height features and multi-frequency height encoding. The system employs a lightweight encoder to integrate geometric and local features to produce semantic BEV maps and object bounding boxes. IndoorBEV has a small parameter count of 0.6M and achieves low latency of 169.6 ms on target hardware. Evaluations confirm a favorable balance between perception accuracy, speed, and memory consumption.
DexPolicy introduces scheduled exploration to guide trajectory-guided dexterous manipulation by explicitly controlling the exploration scale during training. It anneals the exploration from broad to narrow as training progresses while keeping other parameters fixed. Experiments across different policy-optimization settings, including PPO, GRPO, and FPO, showed improvements in deterministic target success. Specifically, the FPO method achieved a higher mean target success than linear decay in some evaluations. The study concludes that schedules should be judged by terminal task success under intended execution conditions.