Can a Robot Hand Be Both Precise and Safe at the Same Time? Two New Papers Say Yes, With Caveats

Model Predictive Control (MPC) addresses this by optimising over a receding horizon, predicting future states and solving a constrained optimisation problem at each timestep. The challenge is that MPC is computationally expensive, and dexterous control requires running at hundreds of hertz. The trick the authors use in both papers is a feedforward linearisation that reduces the plant, regardless of the underlying actuator type (hydraulic, cable-driven, pneumatic, twisted-string, or series-elastic), to a constant-coefficient double integrator. Because the state matrix is constant, the quadratic program cost inverse can be precomputed offline. At runtime, you are solving a much smaller problem. The first paper reports running a 10-step receding-horizon QP at 500 Hz. The second paper claims 1 kHz or above for the arm end-effector layer.

An augmented Kalman filter estimates disturbance states from encoder measurements alone, without force-torque sensors, and drives steady-state error to zero under any constant contact load. That is the offset-free part of the architecture, and it is worth noting that this class of offset-free MPC is not new in general. The authors explicitly ground their work in the constant-$A_d$ offset-free MPC literature and in prior pHRI-specific work. What they are claiming as novel is the extension of that architecture to dexterous fingers (with actuator-agnostic feedforward) and to floating-base whole-body control.

The numbers

The headline result from the finger paper is striking. On a hydraulically actuated finger, the 500 Hz Kalman MPC achieves 0.5 mrad RMS tracking error, 0.1 mrad steady-state error, and 6.6 mrad peak deflection under a 1.5 Nm contact load. The authors compare these figures to classical impedance control and report improvements of 183x, 1500x, and 23x respectively.

Those multipliers are large enough to invite scrutiny. A 1500x improvement in steady-state error is, to be precise, not a marginal gain over a baseline; it is a qualitative shift. The explanation is straightforward in theory: classical impedance accumulates steady-state error by design, while the Kalman disturbance estimator drives it to zero. The 1500x figure is therefore somewhat comparing a method that was never designed to eliminate steady-state error against one that explicitly does. I know I am being picky here, but the framing matters for understanding what is genuinely new versus what is a favourable baseline choice.

The effective stiffness range reported, 18 to 323 Nm/rad depending on update rate, is independently verified within the paper. The authors also demonstrate the architecture on a 16-DOF LEAP Hand in MuJoCo simulation, where the system recovers from 2.5 N grasp-load disturbances within 0.7 seconds.

For the whole-body paper, the architecture becomes a three-level stack. A centroidal MPC plans contact forces over a 500 ms horizon. A priority-driven whole-body control (WBC) layer resolves balance into joint torques using contact-consistent null-space projection. The residual null space is then governed by the receding-horizon QP that handles pHRI disturbances. The authors prove an Impedance Equivalence Theorem showing that the infinite-horizon limit of their controller recovers a classical task-space impedance law, but with effective mass, damping, and stiffness that adapt to posture and contact configuration. Validation is reported on a 17-DOF biped and on the Unitree G1 humanoid, both in simulation.

That last detail matters. Both validations are simulation-only for the whole-body case.

What is genuinely new, and what is incremental

This is where I want to be careful, because the papers sit in a dense literature and the novelty claims require unpacking.

The offset-free MPC architecture with Kalman disturbance augmentation is established. The constant-$A_d$ formulation the authors reference is prior work, and they are explicit about this, which is methodologically honest. The stability, recursive feasibility, and input-to-state-stability guarantees they claim are inherited by preserving that architecture's assumptions, not independently derived. That is a legitimate engineering contribution, extending a proven framework to a new domain, but it is incremental over the theoretical foundations.

What appears to be genuinely new, or at least a meaningful synthesis not previously demonstrated together, is the combination of: actuator-agnostic feedforward linearisation that handles the full range of tendon transmission types under a single constant-coefficient model; the extension of this to dexterous multi-finger hands with hard constraints on contact force (specifically citing ISO/TS 15066, the technical specification for collaborative robot safety), actuation limits, and jerk; and the three-level extension to floating-base platforms with a contact-mode-switching protocol that preserves disturbance estimate continuity across mode transitions via covariance inflation.

Actually, the research shows that the covariance-inflation protocol for contact-mode switching is probably the most interesting individual contribution in the second paper. Contact-mode switches are a persistent challenge for any estimator-based controller because they represent discontinuities in the system model. The approach here, inflating the Kalman filter covariance at switch time to acknowledge increased uncertainty rather than resetting the estimate, is conceptually clean. Whether it is robust across the full range of real-world contact transitions remains unclear from the simulation results alone.

The encoder-only disturbance estimation is also worth highlighting. Force-torque sensors are expensive, fragile, and add mechanical complexity to hands and wrists. A controller that achieves zero steady-state error without them, using only joint encoders, has practical value independent of the MPC framing.

Open questions

Several things I would want to see before updating strongly on these results.

First, the whole-body validation is entirely simulation-based. Simulation results on MuJoCo with a Unitree G1 model are useful for validating the control architecture's logical consistency, but the gap between simulation and hardware for humanoid whole-body control is well-documented and not trivial. Contact dynamics in simulation are typically smoother and more predictable than on real hardware. The 1 kHz operation claim for the arm layer needs hardware verification.

Second, the finger paper does include hardware results, which is the stronger of the two contributions in terms of empirical grounding. But the platform is a single hydraulically actuated finger. The paper also presents a 16-DOF LEAP Hand simulation, but the hardware demonstration is one finger under one contact load condition (1.5 Nm). It is too early to say how the framework performs across the full range of manipulation tasks, varying contact geometries, and dynamic loads that a real dexterous hand encounters.

Third, the sample of contact conditions tested is small. This is not a criticism unique to these papers; it is a structural limitation of laboratory robotics research. But the 183x and 1500x improvement figures are derived from a specific experimental setup, and they have not been replicated by independent groups.

Fourth, the actuator-agnostic claim deserves scrutiny. The feedforward linearisation is derived algebraically and the authors argue it applies across hydraulic, cable, pneumatic, twisted-string, and series-elastic actuators. The worked example platform is hydraulic. Whether the linearisation holds in practice for cable-driven systems with significant hysteresis, or for twisted-string actuators with nonlinear transmission ratios at large deflections, is not empirically demonstrated here. The algebraic argument is sound given its assumptions; the assumptions themselves are worth examining.

Fifth, and this is something the papers do not fully address, the ISO/TS 15066 force constraints are enforced as hard constraints in the QP. That is appropriate for collaborative robot safety. But the constraint tightness interacts with the feasibility guarantees in ways that depend on the specific constraint margins chosen. The papers report recursive feasibility in the theoretical framework, but a more complete treatment would characterise how close to the constraint boundaries the controller operates during typical manipulation tasks.

This raises questions about deployment readiness, well, multiple things really, including certification pathways for safety-critical constraints enforced via MPC in commercial humanoid platforms.

The broader context is that the field is genuinely hungry for controllers that handle the precision-compliance tradeoff without requiring force-torque sensors or sacrificing real-time performance. The approach here is technically serious and grounded in a well-understood theoretical framework. The finger hardware results are encouraging. The whole-body results need hardware validation before strong claims about humanoid pHRI can be made.

For researchers working on dexterous manipulation or humanoid control, both preprints are worth reading carefully. For anyone hoping these papers represent a solved problem, the evidence does not support that reading yet. What they represent is a coherent, theoretically grounded framework with one solid hardware demonstration and promising simulation results. That is a meaningful contribution. It is also, to be precise, exactly where the hard work begins.

Mark Kowalski · 19 hours ago · 6 min