Contact control is hard because the obvious solution to slip is often more force, and more force can damage the object. A useful gripper controller has to reduce slip without simply overgripping.

The nmGrip line treats this as a residual-control problem. Keep a commercial-like impedance controller as the baseline. Add a predictive bio-grip HRR layer to reduce the remaining slip and damage-risk signals. Then test whether more memory geometry, such as NdFractal, improves the tradeoff.

The latest v33 NdFractal experiment did not produce a simple “bigger model wins” result. In stationary 100k-step validation, the commercial-like control RMS was 0.382718. The v32 predictive bio-grip HRR model lowered RMS to 0.254460. The v33 NdFractal GPU1024 variant lowered it further to 0.248397, a 2.38% improvement over v32 and 35.10% better than control, but with a 2.60% worse overgrip measure than v32.

Under regime shift, v33 GPU1024 improved RMS by 2.62% over v32 and 32.20% over control, again with worse overgrip. The cpu64+gpu1024 v33 variant slightly worsened RMS but improved overgrip relative to v32.

That is exactly the kind of result a contact product should disclose. There is no single scalar win. There is a tradeoff between slip, overgrip, runtime, and regime behavior.

The commercial version should let a customer choose the operating point. Fragile objects may prefer less overgrip. Harsh industrial picks may prefer slip reduction. The value of TightLoop Grip is not only lower RMS; it is making that tradeoff explicit and testable.