LIFT (Reactive Force Injection for VLA post-training), proposed by Shanghai Jiao Tong University's Lu Ce-Wu and Wen Chuan teams together with Noematrix, addresses a blind spot of vision-language-action (VLA) models: contact-rich moments where vision barely changes. Accepted to CoRL 2026, code open-sourced.
The problem: force data, not algorithms
Pretrained VLAs rely heavily on vision. Once the robot makes contact — did the book hit its neighbor, is the push direction off, is it fully inserted — visual observations are nearly uninformative, and occlusion, depth ambiguity, and small force errors push execution off the offline demonstration distribution. Force/torque data is tightly bound to specific robot hardware and end-effectors, expensive to collect, and hard to scale via handheld devices like UMI or first-person video. Adding a force pathway also risks degrading pretrained knowledge.
Three design choices that preserve the pretrained prior
LIFT attaches a reactive action expert beside the original action expert rather than modifying it:
1. Weight copying — the reactive expert is initialized with the pretrained action expert's weights; the original action flow keeps its intra-block full attention untouched. 2. Zero-initialized cross-attention — force memory is injected as a residual; the output projection is zero-initialized, so at step 0 force contributes exactly zero update, and pretrained outputs change only as post-training learns nonzero residuals. 3. Shifted causal attention mask — reactive tokens attend to the vision-language prefix, subsequent original action tokens, and their own causal prefix, so their context matches the original full-attention expert.
Slow path and fast path
With action chunk length H=10 at 10 Hz, each chunk covers one second. The vision-language prefix KV cache is computed once per chunk (~1 Hz, slow context). Within each chunk, every control step reads the latest force frame, encodes causal force memory, and refreshes reactive actions at 10 Hz. Force memory uses a single-layer causal GRU (hidden width 512, output width 1024) over the six-axis F/T sequence; raw sensor data at >1000 Hz is downsampled to 10 Hz. A delay-alignment causal mask (offset L=3) blocks force measurements not yet available at inference time.
Using 20-30 online episodes
Each online session lasts ~2-3 hours and collects 20-30 episodes, but only human-intervention frames enter the correction set (operators take over on policy stalls, force-trajectory safety threshold violations, or overlong episodes). Every 100 training steps, a new checkpoint is pushed to the inference server, closing the collect → post-train → deploy loop. Training uses a fixed 1:1 offline:online sampling ratio; vision-only batches use zero-force placeholders with force memory masked before cross-attention (no gradient to the force encoder), while online correction batches keep real force memory.
Benchmark results (vs. vision-only π0.5 with online DAgger)
| Task | Vision-only online DAgger | LIFT | Scoring | |---|---|---|---| | Towel folding | 73.3 | 84.2 | staged; full score for second fold | | Book insertion | 36.7 | 58.3 | collision-free insertion + fully pushed | | Hanoi ring placement | 26.7 | 56.7 | binary success rate |
Paper-table peak scores: book insertion baseline peak 0.4 vs LIFT 0.6; Hanoi baseline peak 0.3, final 0.2, LIFT peak and final 0.6. Each checkpoint is evaluated with 3 sets of 30 autonomous rollouts.
A counterintuitive ablation: removing reactive force memory and feeding single-frame force raises towel folding peak to 0.95, but book insertion drops to 0.4 and Hanoi peaks at 0.5 then falls to 0.1, 0, 0.1. Tasks requiring contact-phase estimation need a short force history; a single frame is not enough.
Platform and stated limitations
Hardware: single Flexiv Rizon 4S arm with Robotiq 2F-85 gripper, position control at 10 Hz, a wrist-mounted iPhone providing 10 Hz monocular RGB. Teleoperated demonstrations via Flexiv TDK bilateral master-slave. Training: single node, 8× A800, based on openpi with π0.5 backbone.
Limitations per the paper: human intervention limits data-collection efficiency and burdens operators; takeover timing is hard to judge and affects correction quality; all experiments are single-arm only.
Why it matters
Most embodied-AI work adds data, simulation, or world models upstream. LIFT works downstream: pretrained models provide the prior, and post-training adds only a high-frequency, low-dimension, hardware-specific modality — teaching contact behavior with just 20-30 online episodes. The authors argue the paradigm applies to any MoT-based pretrained model, not just VLAs. For engineers, the key signal is the slow/fast separation: expensive vision-language context computed once per second, cheap force feedback computed every step — no pretrained-model changes or large-scale re-collection required.
References
- Yi Wang et al., "Never Too Late for Force: Accelerating VLA Post-Training with Reactive Force Injection", arXiv:2607.14236v2, CoRL 2026
- Project page: lift-policy.github.io; code: github.com/y-wng/lift