82.9% when we STACKED frames. Why? The action\n# depends on MOTION, and one still frame can't show it. Here's the mechanism.\nr=np.random.default_rng(1); N=500\nposn=r.uniform(0,1,(N,2)); vel=r.uniform(-.1,.1,(N,2)) # the action to predict = velocity\nprev=posn-vel # where it was one frame ago\nfrom numpy.linalg import lstsq\nA1=np.c_[posn,np.ones(N)]; w1,*_=lstsq(A1,vel,rcond=None); e1=float(((A1@w1-vel)**2).mean())\nA2=np.c_[posn,prev,np.ones(N)]; w2,*_=lstsq(A2,vel,rcond=None); e2=float(((A2@w2-vel)**2).mean())\nprint(f\"one frame -> velocity MSE {e1:.3e}\")\nprint(f\"two frames -> velocity MSE {e2:.3e}\")\nassert e2 < 0.2*e1, \"stacking frames should make the motion recoverable\"\nprint(\"PASS - a single frame can't show motion; stacked frames reveal velocity. That is exactly why the real-pixel VLA stacks 3 frames. Now open /research/vla and run the released models.\")","label":"Why one frame isn't enough — stack them"}],"intro":"Compare one-frame and two-frame prediction to see why motion needs memory.","key":"vision-language-action/motion","kind":"python","title":"Motion needs memory"}">