0916_r1lite_ee_gr00t1.7_tcfm_50000

0916_r1lite_ee_gr00t1.7_tcfm_50000

GR00T-N1.7-3B finetuned on R1 Lite end-effector data (260413_r1lite_ee_gr00t, 50 episodes / 72,474 frames, task foldhoodie) with the auxiliary trajectory-consistency losses from arXiv:2605.08511 (Trajectory-Consistent Flow Matching for Robust Visuomotor Policy Learning).

50,000 steps, batch 16, lr 1e-4, 15.6 h on one H200.

Action / state space

EE poses are xyz + rot6d, relative to the current pose (ActionRepresentation.RELATIVE, ActionType.EEF, ActionFormat.XYZ_ROT6D). Grippers are absolute. Action chunk length 32, padded to the model horizon of 40. Video: head_rgb (720p) + left_wrist_rgb / right_wrist_rgb (360p).

Training objective

L = L_single_step + 0.5*L_multistep + 0.1*L_action + 0.1*L_vel

matching the paper's weights (λ_c = 0.5, λ_a = 0.1, λ_v = 0.1). The paper's L_rect is already this model's built-in single-step flow matching loss, so it carries the paper's λ_r = 1.0 implicitly; the paper's Consistency-FM baseline term L_CFM is not implemented here. Flow time is drawn uniformly (noise_time_distribution="uniform"), as L_rect prescribes, rather than from GR00T's default Beta(1.5, 1) skew. Auxiliary weights ramp linearly over the first 2,500 steps.

Counts follow the paper: K = 3 segments of S = 4 Euler steps for L_multistep, S_act = 5 for L_action, S = 5 probes for L_vel.

Final losses

L_vel rises over training, but the field did not get less smooth -- it got bigger. L_vel is an unnormalised squared difference of velocities, so its units are those of v**2. Measuring this checkpoint against an identically initialised one on the same inputs:

L_vel grew 43x while the field's own magnitude grew 82x, so smoothness relative to the field improved by 1.9x. The training log shows the same thing: single_step starts at 1.267, which is the second moment of the target x1 - x0 (the model predicts ~0 at init), and ends at 0.008, so E[v**2] went from ~0 to ~1.27 while L_vel rose only 6.4x.

Open-loop evaluation

10 trajectories x 800 frames, execution_horizon 16, on the training dataset. MAE is the meaningful column — the flat MSE is dominated by the gripper channel, whose range is 0-100 against ~1 for the pose channels.

Use denoising_steps=8 with the default Euler sampler. It beats the shipped 4-step setting on 9/10 trajectories (sign test p = 0.021); going to 16 buys a further 4%, which is inside the noise.

RK4 is not recommended, despite being one of the paper's four contributions. At matched compute it lost on 20 of 20 trajectories (p = 0.002 at each budget): 2.31x worse than Euler at 8 NFE and 1.63x worse at 16 NFE. The paper reports that RK4 fails without L_vel and works with it; L_vel is enabled here at the paper's weight and RK4 still loses. RK4 reads the field at half-grid times and at probe-displaced states, which nothing in training ever supervises.

Caveats

• No auxiliary-loss-off control was trained on this dataset, so the contribution of the auxiliary terms themselves is not established here. The numbers above compare samplers on one checkpoint, not recipes.
• Evaluation is on the training dataset; these are fit numbers, not generalisation numbers.
• Single task (foldhoodie), single embodiment, 50 episodes.

Usage

python gr00t/eval/open_loop_eval.py \
  --model-path XYZPIT/0916_r1lite_ee_gr00t1.7_tcfm_50000 \
  --dataset-path <your lerobot dataset> \
  --embodiment-tag new_embodiment \
  --denoising-steps 8