A Humanoid Controller That Feels What It Grabs — Without Force Sensors

FORGE replaces wrist force/torque sensors with a learned momentum observer, letting humanoid robots squeeze, lift, and walk using only joint-level proprioception and sparse wrist poses.

Imagine you're carrying a heavy grocery bag but wearing thick winter gloves. You can't feel the bag's weight through your fingertips, so instead you infer the load from the strain in your shoulder and elbow — how hard your muscles are working against gravity, how your balance shifts. FORGE does exactly this for a humanoid robot: it reconstructs wrist-level interaction forces not from a dedicated force sensor but from the discrepancy between what the low-level joint controllers commanded and what actually happened. That proprioceptive trick is the paper's core mechanism, and it's the reason the system matters. The committed claim: a single RL policy, trained in simulation with no full-body human motion capture and no joint-angle references, can simultaneously walk, balance, track bilateral wrist targets in SE(3), estimate quasi-static contact forces, and modulate applied squeeze force — all from sparse teleoperation inputs (pelvis-relative wrist poses plus coarse velocity commands). No wrist force/torque sensors at deployment. The paper calls this "force-aware multi-granularity control," which is a mouthful, but the substance is real: a unified controller that handles loco-manipulation with contact awareness, trained without the expensive motion-capture pipelines that most humanoid controllers still depend on. The architecture is a coarse-to-fine RL policy stack. A frozen locomotion prior (pre-trained for stable walking) provides the base; a residual manipulation policy sits on top, modifying only what it needs to. The residual is asymmetric — observation-space in training is richer than at deployment, a standard sim-to-real trick. Force estimation comes from a momentum observer that compares reconstructed PD torques against expected dynamics, producing low-frequency wrist-equivalent force estimates. A quasi-static virtual-target loop then adjusts the wrist setpoint using the estimated force and an optional desired-force input. The whole thing trains in Isaac Gym with contact-force supervision from the simulator. On the ladder: the closest prior art is the family of whole-body humanoid controllers from 2023-2024 (OmniH2O, ExBody, HumanPlus) that track motion-capture trajectories and handle locomotion but don't reason about contact forces. FORGE doesn't claim to beat these on pure locomotion quality — that's not the point. It claims a new capability axis: force-conditioned manipulation during locomotion, without requiring motion-capture demonstrations or wrist F/T sensors. The paper shows stable locomotion under push perturbations, SE(3) wrist tracking with <3 cm position error and <5° orientation error, and quasi-static force estimation within ~5N of ground truth during a box-squeeze task. A downstream diffusion policy executes a squat-and-lift across payload conditions (0-5 kg), demonstrating the controller as a foundation for learned manipulation skills. Integrity is simulation-only with no real-hardware deployment reported — a significant caveat. All force estimation validation is against simulator ground truth, which means the momentum observer has never faced real actuator backlash, joint friction, or sensor noise. The benchmarks are reasonable for a sim paper (tracking error, force estimation error, perturbation recovery), but they're all self-graded. No community benchmark exists for humanoid loco-manipulation force estimation, so the authors built their own evaluation suite. Code and a project website are provided. The milestone to watch is sim-to-real transfer of the force estimation loop. Current humanoid control papers (H1, Unitree G1 deployments) have shown locomotion transfer works; wrist tracking transfer is emerging. But force estimation from proprioception in the real world — where actuator models are imperfect, cable routing creates unmodeled friction, and thermal drift changes joint behavior — is a different beast. If FORGE's momentum observer holds up on real hardware with <10N force estimation error during sustained contact, that's the unlock: sensor-free contact-aware humanoid manipulation becomes practical for logistics and caregiving tasks. The obvious experiment not run is real-robot deployment. The authors have a project website with simulation videos only. My honest read: this is a (c) situation — they're saving hardware results for the next paper or a concurrent submission. The sim framework is mature, the policy architecture is deployment-ready, and the team (Tsinghua/Shanghai AI Lab) has hardware access. A pure (a) budget constraint seems unlikely given the institutional backing. The secondary missing experiment is adversarial perturbation during contact — what happens when someone pushes the robot while it's squeezing a box, forcing the force estimator and balance controller to fight over the same degrees of freedom?