← Paper Digest

TMA: Sim-to-Real Forceful Humanoid Manipulation via Tactile-Driven Motor Adaptation

Yiting Chen, Kelin Yu, Sarthak Kaingade, Bryan Trinh, Kaiyu Hang, Roberto Martín-Martín, Nima Fazeli, X. Alice Wu

November 2026

The G1 grasps a brass round knob, turns it and pushes the door open.

TL;DR A humanoid learns to open doors with round knobs entirely in simulation, adapting its grip online from 2 s of fingertip force history. On a Unitree G1 with four real knobs it opens the door in 27 of 40 trials, against 6 for a reactive tactile policy and none for RMA.

What they do

Opening a door with a round knob is hard for a humanoid hand. The fingers must hold a frictional grasp, build up enough torque to release a spring-loaded latch without slipping, and then push the door while the robot stays balanced. Friction and knob stiffness cannot be seen in vision or proprioception, but they show up in how fingertip forces change over time. TMA (Tactile-Driven Motor Adaptation) trains a policy for this task entirely in simulation and transfers it zero-shot to a Unitree G1 with BrainCo Revo2 hands.

How it works

  • Whole-body split. The policy drives the right arm and hand (13 joints). A frozen, pretrained locomotion controller keeps the legs and waist balanced.
  • Cone-coordinate touch. Each of four fingertip sensors gives a 3D force, encoded as normal load, cone angle and shear direction. A grip gate scales the knob and door rewards and gives full credit only when the thumb and at least two opposing fingers touch the knob with their sensing pads.
  • Motor adaptation from touch. An oracle policy is trained with PPO on privileged knob stiffness, friction and door state. An adaptation encoder then learns to estimate the same latent from 2 s of observation history, so the policy can run without privileged inputs.
  • Randomization of tactile calibration, noise and dropout, and of knob shape, spring stiffness, friction and panel mass.

Results

  • Simulation ablation: full cone-coordinate touch opens the door in 80.9% of episodes, against 73.6% with binary contact, 74.2% with normal load only and 65.1% with shear only.
  • Real world, four commercial knobs, 10 trials each: TMA opens the door in 27 of 40 trials (32 grasps). The reactive tactile policy opens it in 6, and RMA without touch in none, since its hand drifts off the knob.

Limitations

Only one task is shown, and the robot is placed by hand in front of the door. Knob pose comes from an AprilTag, and tape on the knob rim adds friction. Performance dropped markedly once the fingertip pads were damaged.

Videos

Another real-world rollout from reach to door opening.
Opening the door with a silver round knob.

Figures

From Figure 3: Overview of TMA. (a) The manipulation policy commands the right arm and hand while a pretrained, froze...
From Figure 3: Overview of TMA. (a) The manipulation policy commands the right arm and hand while a pretrained, frozen locomotion controller commands the legs and waist. (b) Each fingertip force is encoded in cone coordinates: normal load n, cone angle alpha and the shear direction as sin phi, cos phi. (c) A privileged encoder and the policy are first trained together; then both are frozen and an adaptation encoder learns to estimate the interaction latent from 2 s of observation history. (d) At deployment the adaptation encoder updates the latent online from tactile history.
From Figure 4: Real-world rollouts of TMA, one row per commercial knob: reaching, multi-finger grasping, knob rotatio...
From Figure 4: Real-world rollouts of TMA, one row per commercial knob: reaching, multi-finger grasping, knob rotation and door opening. An AprilTag gives the knob pose, and tape on the knob rim increases friction.
From Figure 6: Real-world results, 10 trials per knob and method. Dark segments are door-open successes, light segmen...
From Figure 6: Real-world results, 10 trials per knob and method. Dark segments are door-open successes, light segments grasp-only successes; labels give door-open over grasp successes.
Abstract

Recent humanoid loco-manipulation in articulated environments has largely focused on geometry-dominant interactions, where success is primarily determined by geometric alignment and motion execution. In contrast, forceful multi-finger manipulation requires the robot to adapt to real-world uncertainties in contact and interaction dynamics. We present Tactile-Driven Motor Adaptation (TMA) and an integrated whole-body system for zero-shot sim-to-real forceful humanoid manipulation, and study round-doorknob opening as a challenging benchmark requiring stable multi-finger frictional grasping, sufficient torque generation without slip, force-regulated pushing, and whole-body stability. TMA structures policy learning around interactions observable through tactile feedback, uses cone geometry as an inductive bias for multi-finger contact representation, and adapts motor behavior online from tactile history. An upper-body manipulation policy is trained alongside a pretrained general locomotion controller to coordinate dexterous interaction and whole-body stabilization. Trained entirely in simulation, TMA transfers zero-shot to a Unitree G1 humanoid with BrainCo Revo2 multi-fingered hands and substantially outperforms standard Rapid Motor Adaptation (RMA) and reactive tactile policy baselines across four commercial round doorknobs in the real world.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.